Seismic fault identification method and system based on noise decoupling and selective retraining
By introducing noise decoupling and selection retraining methods in seismic fault recognition, the coupling problem between teacher model and student model and the problem of unmarked image stability differences are solved, the accuracy and efficiency of seismic fault recognition are improved, and the generalization ability of the model is enhanced.
Patent Information
- Application Number
- CN202411728385.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-02-28
AI Technical Summary
In the prior art, there is a serious coupling problem between the teacher model and the student model, which leads to excessive dependence on the teacher model when learning pseudo-labels, lack of diversity and independence, and the stability differences of unlabeled images and the reliability of pseudo-labels lead to serious deviations in pathological pseudo-labels, which weakens the learning ability and generalization ability of the student model.
A seismic fault recognition method based on noise decoupling and selection retraining is proposed. By adding random noise in the retraining stage and performing selection retraining, the accuracy and efficiency of seismic fault recognition are improved. The teacher model is used to generate and filter the optimal pseudo-labels, and combined with random noise to enhance data diversity, the trained student model is highly generalized and can accurately identify seismic faults, reducing the dependence on a large number of labeled data.
Through the methods of noise decoupling and selection and retraining, the student model achieves a balance between knowledge inheritance and independent learning, improving the accuracy and efficiency of seismic fault recognition, enhancing the generalization ability of the model, and reducing dependence on labeled data.
Smart Images
Figure CN119540565A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to an earthquake fault identification method and system based on noise decoupling and selective retraining. Background Art
[0002] Deep learning fully supervised semantic segmentation models based on a large number of labeled seismic images have made great progress. However, in seismic imaging, pixel-level labels for image semantic segmentation are difficult to obtain, require a lot of time and cost, and require extremely high professionalism.
[0003] Semi-supervised learning has become the dominant seismic data interpretation strategy, which learns fault recognition models from a limited number of labeled images and an arbitrary number of unlabeled images. Among them, self-training is the simplest pseudo-labeling and semi-supervised learning strategy, where the retrained student model is supervised by pseudo-labels, which are predicted by the teacher model, which is trained on labeled images.
[0004] However, in the context of seismic exploration, the following problems still exist: On the one hand, there is a serious coupling problem between the teacher model and the student model, which causes the student model to overly rely on the output of the teacher model when learning the pseudo-labels of the teacher model in a supervised manner, lacking diversity and independence, thereby limiting the student model's ability to learn additional semantic information. On the other hand, the stability differences of different unlabeled images and the reliability of the corresponding pseudo-labels cause the pathological pseudo-labels to introduce serious deviations in the iterative optimization process of the model, further weakening the learning and generalization capabilities of the student model. Summary of the invention
[0005] In order to solve the above problems, the present invention proposes an earthquake fault identification method and system based on noise decoupling and selective retraining. By adding random noise and performing selective retraining in the retraining stage, the accuracy and efficiency of earthquake fault identification are improved. The teacher model is used to generate and screen the optimal pseudo-labels, and random noise is combined to enhance data diversity. The trained student model has strong generalization ability, can accurately identify earthquake faults, and reduces the dependence on a large amount of labeled data.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] In a first aspect, the present invention provides a method for earthquake fault identification based on noise decoupling and selective retraining, comprising:
[0008] Get labeled images and unlabeled images;
[0009] Using the labeled image to train a teacher model, saving the current optimal teacher model based on a preset period, and generating pseudo labels for unlabeled images; selecting the optimal pseudo label from the pseudo labels generated by different optimal teacher models based on the F1 score;
[0010] Add random noise to the unlabeled image corresponding to the optimal pseudo-label and the remaining unlabeled images to obtain two groups of images; train the student model with the labeled images and the optimal pseudo-label images after adding noise, and generate pseudo-labels for the other group of noisy images; combine all images to train the student model until the training is completed;
[0011] The original image to be classified is input into the trained student model to classify and identify earthquake faults.
[0012] Preferably, the method of saving the current optimal teacher model based on a preset period is, specifically, during the pre-training process, saving a current optimal teacher model as the optimal teacher model every preset training period.
[0013] Preferably, the method of selecting the optimal pseudo label from the pseudo labels generated by different optimal teacher models based on the F1 score is specifically as follows:
[0014] Saving the current best optimal teacher model based on a preset period, and using the optimal teacher model to generate pseudo labels for unlabeled images;
[0015] Calculate the F1 scores of the pseudo labels generated by different training cycles, and select the optimal pseudo labels based on the highest F1 score.
[0016] Preferably, random noise is added to the unlabeled image corresponding to the optimal pseudo-label and the remaining unlabeled images respectively to obtain two groups of images, specifically: random noise is added to the unlabeled image corresponding to the optimal pseudo-label to obtain a first noisy image; random noise is added to the remaining unlabeled images to obtain a second noisy image.
[0017] Preferably, the method of training the student model with the labeled image and the optimal pseudo-label image after noise addition, and generating pseudo-labels for another set of noisy images, is specifically as follows:
[0018] A student model is trained using the labeled image and the first noisy image, and a pseudo label is generated for the second noisy image using the student model.
[0019] Preferably, the combining of all images to train the student model until the training is completed is specifically: using the labeled image, the first noisy image and the second noisy image to train the student model until a trained student model is obtained.
[0020] In a second aspect, the present invention provides an earthquake fault identification system based on noise decoupling and selective retraining, comprising:
[0021] A data acquisition module, used for acquiring labeled images and unlabeled images;
[0022] A teacher model training module is used to train a teacher model using the labeled image, save the current optimal teacher model based on a preset period, and generate pseudo labels for unlabeled images; and select the optimal pseudo label from the pseudo labels generated by different optimal teacher models based on the F1 score;
[0023] The student model training module is used to add random noise to the unlabeled image corresponding to the optimal pseudo-label and the remaining unlabeled images to obtain two groups of images; the student model is trained with the labeled image and the optimal pseudo-label image after adding noise, and pseudo-labels are generated for another group of noisy images; the student model is trained with all images until the training is completed;
[0024] The fault identification module is used to input the original image to be classified into the trained student model to classify and identify earthquake faults.
[0025] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the earthquake fault identification method based on noise decoupling and selective retraining described in the first aspect.
[0026] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the earthquake fault identification method based on noise decoupling and selective retraining described in the first aspect are implemented.
[0027] Compared with the prior art, the present invention has the following beneficial effects:
[0028] The present invention proposes an earthquake fault identification method and system based on noise decoupling and selective retraining. First, a self-training strategy Fault-NE based on random noise decoupling is proposed, which adds random noise to unlabeled images to expand data and decouple teacher and student models. Secondly, a selective retraining strategy Fault-NE+ is proposed, which selects and trains reliable pseudo-labels based on model prediction stability; the teacher model is trained based on labeled images and unlabeled images, and the student model is trained based on labeled images and unlabeled images with random noise added, which can balance the student model between knowledge inheritance and autonomous learning. In addition, the proposed composite loss function can guide the optimization model training to learn pixel-level feature representations that are intra-class affinities and inter-class separables. Using the trained student model, key features can be effectively learned, thereby improving the accuracy of earthquake fault identification.
[0029] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present invention but do not constitute a limitation of the present invention.
[0031] Figure 1 A main flow chart of an earthquake fault identification method based on noise decoupling and selective retraining provided by an embodiment of the present invention;
[0032] Figure 2 A schematic diagram of a self-training strategy Fault-NE based on noise decoupling provided in an embodiment of the present invention;
[0033] Figure 3 A schematic diagram of the Fault-NE+ retraining strategy selected according to an embodiment of the present invention;
[0034] Figure 4 A schematic diagram of a semi-supervised learning framework provided by an embodiment of the present invention;
[0035] Figure 5 A schematic diagram showing a comparison of visualization results of the Fault-NE strategy, the Fault-NE+ strategy and other training strategies on a synthetic data set provided by an embodiment of the present invention;
[0036] Figure 6 A schematic diagram showing a comparison of the visualization results of the Fault-NE+ strategy provided in an embodiment of the present invention and other training strategies on a field data set. DETAILED DESCRIPTION
[0037] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0038] Embodiment 1
[0039] like Figure 1 As shown, this embodiment discloses a method for identifying earthquake faults based on noise decoupling and selective retraining, comprising the following steps:
[0040] S1: Obtain labeled images and unlabeled images;
[0041] S2: training a teacher model using the labeled image, saving the current optimal teacher model based on a preset period, and generating pseudo labels for unlabeled images; and selecting the optimal pseudo label from the pseudo labels generated by different optimal teacher models based on the F1 score;
[0042] S3: Add random noise to the unlabeled image corresponding to the optimal pseudo-label and the remaining unlabeled images to obtain two groups of images; train the student model with the labeled images and the optimal pseudo-label images after adding noise, and generate pseudo-labels for the other group of noisy images; train the student model with all the images until the training is completed;
[0043] S4: Inputting the original image to be classified into the trained student model to classify and identify earthquake faults.
[0044] Next, combine Figure 1 , a method for earthquake fault identification based on noise decoupling and selective retraining disclosed in this embodiment is described in detail.
[0045] (I) Self-training strategy based on random noise decoupling Fault-NE
[0046] To address the problems of overfitting of noisy labels and coupling of teacher-student model prediction, a self-training strategy Fault-NE based on random noise decoupling is proposed in the retraining stage, such as Figure 2 As shown in the figure, the purpose is to break the direct dependence between models and enable the student model to learn richer semantic information. By adding an appropriate amount of random noise to the unlabeled images, a more challenging optimization goal is proposed for the student model, and the prediction decoupling between the teacher model and the student model is achieved, thus enhancing the learning ability of the model.
[0047] Specifically, given a data set X = {X L ,X U},in is a subset of labeled images, is a subset of unlabeled images, x i is the input image, y i For its corresponding label, L and U are the number of labeled images and unlabeled images respectively. Generally, L is much smaller than U.
[0048] In the pre-training phase, on the labeled dataset X L Train the teacher model T on the network and use the supervised loss function L S renew.
[0049] In the retraining phase, the teacher model T is used to train the unlabeled data X U Mark and obtain a pseudo-labeled dataset And combine it with the labeled dataset to get a new dataset
[0050] In the retraining phase, Add random noise Noisy(x k ,y k), and retrain the student model S using the noisy dataset, using the unsupervised loss function L ST renew.
[0051] The self-training strategy Fault-NE based on random noise decoupling proposed in this embodiment adds random noise to the unlabeled image during the retraining phase, so that the feature space of the image changes. For the teacher model and the student model, the inputs they face are no longer exactly the same, thereby achieving prediction decoupling between the teacher model and the student model, so that the student model can generate high-quality pseudo labels.
[0052] (II) Selecting the retraining strategy Fault-NE+
[0053] In the Fault-NE strategy, since the inherent reliability and stability of individual samples are not distinguished, all unlabeled images are treated equally. However, different unlabeled images cannot be equally stable, and the corresponding pseudo-labels cannot be equally reliable. Pathological pseudo-labels will lead to serious bias and potential performance degradation when iteratively optimizing the model. Therefore, this embodiment further proposes to select the retraining strategy Fault-NE+, such as Figure 3 shown.
[0054] By prioritizing reliable and stable unlabeled images for retraining, we fully utilize the entire unlabeled dataset in an easy-to-difficult manner, aiming to evaluate the stability of unlabeled images using a single trained model without the need to manually select a confidence threshold.
[0055] Considering that there is a positive correlation between the segmentation performance of the model and the reliability of the generated pseudo-labels, more reliable pseudo-labels are screened out according to the stability changes of unlabeled images during model training.
[0056] Specifically, in the labeled dataset X L Train the teacher model T and save the optimal teacher model every 30 Epochs during the supervised training process use For unlabeled data Mark and generate pseudo label M ij =T j (x i ).
[0057] Since the training model tends to converge and reaches the best performance in the later stage of training, the reliability of pseudo labels is evaluated by evaluating the F1-Score between the early Epoch pseudo labels and the later Epoch pseudo labels:
[0058]
[0059] Where M ijis the early Epoch pseudo-label, M iK is the pseudo label of the later Epoch, s i is the stability evaluation value.
[0060] After obtaining the stability evaluation values of all unlabeled images, the entire unlabeled dataset is sorted according to the evaluation values, and the top D images with the highest evaluation values are selected to form a new dataset X U1 , new dataset X U1 The corresponding pseudo-label is the optimal pseudo-label.
[0061] Furthermore, in order to decouple the teacher model and the student model, before training the student model, the unlabeled data X is trained based on the Fault-NE strategy. U Add random noise to get the first noisy image X' U1 ; is the remaining unreliable image X U2 =X U -X U1 Add random noise to get the second noisy image X' U2 .
[0062] Afterwards, label the image X with L And the first noisy image trains the student model, that is, in the dataset X L ∪X' U1 The first round of retraining student model S is performed on 1 .
[0063] Use the optimized student model S 1 For the remaining second noisy image X' U2 Re-label and generate pseudo labels.
[0064] In X L ∪X' U1 ∪X' U2 Perform a second round of retraining on the student model S 2 , the training is completed.
[0065] The selection retraining strategy Fault-NE+ proposed in this embodiment evaluates the stability of model prediction, selects reliable pseudo labels, and retrains the reliable labels to improve the model's trust in the pseudo labels.
[0066] (III) Composite loss function
[0067] like Figure 4 As shown in the figure, based on pseudo-label self-training, a composite loss function is proposed to optimize the pixel-level feature representation. By optimizing the distribution of samples in the feature space, pixels with the same semantic labels are encouraged to be close to each other in the feature space, while pixels with different semantic labels are encouraged to be far away from each other. The purpose is to learn feature representations with good intra-class affinity and inter-class separability.
[0068] Specifically, the basic network B θ It is a shared feature extraction network consisting of two branches, one of which is the teacher model G ξ , the other branch is the student model H φ .
[0069] In the labeled image X L By minimizing the supervision loss L S Train the teacher model G ξ :
[0070]
[0071] In the formula, the supervision loss L S is the Focal loss, (θ,ξ) are random initialization parameters, Y is the manual label, and Y L is the segmentation map generated by the teacher model.
[0072] In the labeled image X L and unlabeled image X U The student model H is tested on the combined dataset φ Retraining, using unsupervised loss L ST Update, unsupervised loss L ST It consists of two parts:
[0073] L ST (θ,ξ,φ)=μL STL +(1-μ)L STU (3)
[0074] in,
[0075]
[0076] Where T and S map the input X to the output space, H(·) is the entropy minimization loss between the teacher model T and the student model S, and Y L Generated by the teacher model T prediction, Generated by the student model S prediction.
[0077]
[0078] Where Noisy(·) adds random noise to the original unlabeled image, Y U Generated by the teacher model T prediction, Generated by the student model S prediction.
[0079] In summary, the composite loss function that guides model training is:
[0080] L Total=(1-α)L S (θ,ξ)+αL ST (θ,ξ,φ) (6)
[0081] Where, L ST (θ,ξ,φ) is L STL Loss and L STU The integration of losses, the regularization parameter α is used to balance the supervised loss and the unsupervised loss, and is dynamically adjusted by the Gaussian increasing function:
[0082]
[0083] In the formula, α m is the maximum value of the weighted function, t represents the current iteration Epoch, t m is the maximum Epoch in training.
[0084] This embodiment proposes a composite loss function to learn pixel-level feature representations with intra-class affinity and inter-class separability, thereby enhancing the model's ability to capture details.
[0085] In this embodiment, the teacher model is trained on both labeled images and unlabeled images, while the student model is trained on labeled images and unlabeled images with random noise added. This training method allows the teacher model and the student model to have more overlapping parts in the training data. Since both the teacher model and the student model process labeled images, to a certain extent, the student model can more closely follow the knowledge guidance of the teacher model during the learning process; for unlabeled images, the teacher model training can provide a stable knowledge base for the student model, so that the student model can have a more reliable reference when processing unlabeled images with noise added, reducing the instability of learning caused by noise interference, and at the same time, it can form its own independent feature learning by processing unlabeled images with noise added, thereby achieving a better balance between knowledge inheritance and autonomous learning, and avoiding over-dependence or over-disconnection.
[0086] Figure 5 The figure is a comparison diagram of the visualization results of the Fault-NE strategy, Fault-NE+ strategy and other training strategies on the synthetic dataset. Fault-NE+ achieves the most superior fault segmentation results, showing more complete, accurate and smooth fault lines, which shows that by selecting the retraining strategy, the model's ability to utilize unlabeled images has been significantly improved, ultimately achieving more reliable fault identification.
[0087] Figure 6The diagram below is a comparison of the visualization results of the Fault-NE+ strategy and other training strategies on the field data set. The model trained based on the Fault-NE+ strategy almost completely captures all faults, and the fault locations are very close to the real faults, especially in complex seismic images, where the fault edges are clearer and more coherent, and the fault intersections are more delicately depicted.
[0088] The pixel-level labels for current image semantic segmentation are difficult to obtain, which requires a lot of time and cost. In view of the above problems, the present invention proposes a method and system for earthquake fault identification based on noise decoupling and selective retraining. First, a self-training strategy Fault-NE based on random noise decoupling is proposed. Random noise is added to unlabeled images. Although random noise will deteriorate the clean data distribution, it has a positive impact on unlabeled images, realizes data expansion, and realizes the decoupling of teacher models and student models. Then, in order to alleviate the potential model performance degradation caused by the iteration of pathological pseudo-labels, a selective retraining strategy Fault-NE+ is further proposed. Reliable pseudo-labels are prioritized based on model prediction stability, and selective retraining is performed. Finally, in order to learn pixel-level feature representations with intra-class affinity and inter-class separability, a composite loss function is proposed to guide and optimize model training. The teacher and student models are trained in a specific manner provided in this embodiment, which can balance the student model between knowledge inheritance and autonomous learning.
[0089] Embodiment 2
[0090] This embodiment provides an earthquake fault identification system based on noise decoupling and selective retraining, including:
[0091] A data acquisition module, used for acquiring labeled images and unlabeled images;
[0092] A teacher model training module is used to train a teacher model using the labeled image, save the current optimal teacher model based on a preset period, and generate pseudo labels for unlabeled images; and select the optimal pseudo label from the pseudo labels generated by different optimal teacher models based on the F1 score;
[0093] The student model training module is used to add random noise to the unlabeled image corresponding to the optimal pseudo-label and the remaining unlabeled images to obtain two groups of images; the student model is trained with the labeled image and the optimal pseudo-label image after adding noise, and pseudo-labels are generated for another group of noisy images; the student model is trained with all images until the training is completed;
[0094] The fault identification module is used to input the original image to be classified into the trained student model to classify and identify earthquake faults.
[0095] Embodiment 3
[0096] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the earthquake fault identification method based on noise decoupling and selective retraining as described in the first embodiment above are implemented.
[0097] Embodiment 4
[0098] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the earthquake fault identification method based on noise decoupling and selective retraining as described in the first embodiment are implemented.
[0099] The steps or modules involved in the above embodiments 2 to 4 correspond to those in embodiment 1. For the specific implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.
[0100] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for earthquake fault identification based on noise decoupling and selective retraining, characterized in that: include: Get labeled images and unlabeled images; Using the labeled images to train a teacher model, saving the current best optimal teacher model based on a preset period, and generating pseudo labels for unlabeled images; Filter the best pseudo labels from the pseudo labels generated by different optimal teacher models based on the F1 score; Add random noise to the unlabeled image corresponding to the optimal pseudo-label and the remaining unlabeled images to obtain two groups of images; train the student model with the labeled images and the optimal pseudo-label images after adding noise, and generate pseudo-labels for the other group of noisy images; combine all images to train the student model until the training is completed; The original image to be classified is input into the trained student model to classify and identify earthquake faults.
2. The earthquake fault identification method based on noise decoupling and selective retraining according to claim 1, characterized in that: The method of saving the current optimal teacher model based on a preset period is specifically, during the pre-training process, saving a current optimal teacher model as the optimal teacher model every preset training period.
3. The earthquake fault identification method based on noise decoupling and selective retraining according to claim 1, characterized in that: The method of selecting the optimal pseudo-label from the pseudo-labels generated by different optimal teacher models based on the F1 score is specifically as follows: Saving the current best optimal teacher model based on a preset period, and using the optimal teacher model to generate pseudo labels for unlabeled images; Calculate the F1 scores of the pseudo labels generated by different training cycles, and select the optimal pseudo labels based on the highest F1 score.
4. The earthquake fault identification method based on noise decoupling and selective retraining according to claim 1, characterized in that: The method adds random noise to the unlabeled image corresponding to the optimal pseudo-label and the remaining unlabeled images respectively to obtain two groups of images, specifically: adding random noise to the unlabeled image corresponding to the optimal pseudo-label to obtain a first noisy image; adding random noise to the remaining unlabeled images to obtain a second noisy image.
5. The earthquake fault identification method based on noise decoupling and selective retraining according to claim 4, characterized in that: The method uses the labeled image and the optimal pseudo-label image after noise to train the student model, and generates pseudo-labels for another group of noisy images, specifically: uses the labeled image and the first noisy image to train the student model, and uses the student model to generate pseudo-labels for the second noisy image.
6. The earthquake fault identification method based on noise decoupling and selective retraining according to claim 4, characterized in that: The student model is trained by combining all the images until the training is completed, specifically: the student model is trained by using the labeled image, the first noisy image and the second noisy image until a trained student model is obtained.
7. An earthquake fault identification system based on noise decoupling and selective retraining, characterized in that: include: A data acquisition module, used for acquiring labeled images and unlabeled images; A teacher model training module, used to train a teacher model using the labeled image, save the current optimal teacher model based on a preset period, and generate pseudo labels for unlabeled images; Filter the best pseudo labels from the pseudo labels generated by different optimal teacher models based on the F1 score; The student model training module is used to add random noise to the unlabeled image corresponding to the optimal pseudo-label and the remaining unlabeled images to obtain two groups of images; the student model is trained with the labeled image and the optimal pseudo-label image after adding noise, and pseudo-labels are generated for another group of noisy images; the student model is trained with all images until the training is completed; The fault identification module is used to input the original image to be classified into the trained student model to classify and identify earthquake faults.
8. An earthquake fault identification system based on noise decoupling and selective retraining as claimed in claim 7, characterized in that: The method of saving the current optimal teacher model based on a preset period is specifically, during the pre-training process, saving a current optimal teacher model as the optimal teacher model every preset training period.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the earthquake fault identification method based on noise decoupling and selective retraining as described in any one of claims 1 to 7 are implemented.
10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the earthquake fault identification method based on noise decoupling and selective retraining as described in any one of claims 1 to 7 are implemented.