Transfer learning joint inversion method and system

Abstract

According to the transfer learning joint inversion method and system provided by the invention, cleaning and gradient standardization processing are carried out on actual data, a stable high-quality actual data set is constructed, and for actual detection data without labels, the real detection data without labels can be subjected to gradient standardization processing; a cross-domain feature alignment method of a network end and an output end based on synthetic data and actual data is provided, a transfer learning effect between different data domains is tested by constructing a disturbance learning mode of matching different advanced prediction methods with training, and weak supervision transfer learning is performed on a small amount of manually labeled actual data. And parameter learning of the network in actual data is further optimized through model fine tuning.

Description

Technical Field

[0001] The present invention belongs to the field of geophysical exploration, and in particular relates to a transfer learning joint inversion method and system. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] As tunnel construction gradually shifts to complex geological areas such as high altitudes and deep burial depths, the geological conditions for engineering construction are becoming increasingly complex. Multiple types of disasters such as disaster-causing water bodies, fault fracture zones, strong karst, and weathering deep troughs are coupled and chained. It is difficult to accurately characterize complex geological bodies using only single geophysical methods such as seismic wave method, resistivity method, and electromagnetic method. In the face of high-risk sections under complex geology, it is necessary to use a variety of geophysical methods to conduct comprehensive assessments from multiple angles and dimensions. However, the traditional comprehensive prediction technology based on manual comparative analysis cannot fully utilize the essential characteristics of different geophysical methods, as well as their mutual constraints and data fusion. It is difficult to accurately reveal the characteristics and correlations of the massive information of multiple methods, which affects the accuracy of the detection results.

[0004] In recent years, deep learning algorithms have made significant progress in multimodal data fusion and heterogeneous data fusion with their powerful nonlinear mapping and high-dimensional optimization capabilities. They can effectively mine and analyze the complex mapping relationship between multimodal heterogeneous input data and multi-task output data, and provide new solutions for the joint inversion of multi-source data. However, the accuracy of deep learning methods is extremely dependent on the sample size of training data, and the small amount of labeled sample data has seriously hindered the application of deep learning in the field of tunnel advance prediction.

[0005] Transfer learning is a machine learning method that uses the knowledge of a model trained in one field to improve the performance of tasks in another related field.

[0006] At present, in the field of geophysics, the research on exploration data processing and its application methods based on transfer learning has made great progress. In the field of seismic exploration, transfer learning has been applied to data denoising, inversion imaging, geological structure interpretation and many other aspects; in the field of electrical method, there are also studies on unsupervised inversion methods of DC resistivity based on transfer learning for actual electrical data, and these studies have achieved good results. However, the application of transfer learning in the field of geophysics still has problems such as small amount of data, limited labeled samples and poor labeling quality. How to use a small sample actual data set containing a small amount of valid data to update network parameters is an urgent problem to be solved.

[0007] There are still two challenges in realizing the transfer learning joint inversion method and system that can be applied to tunnel construction: First, there are significant differences between the characteristics of detection data and synthetic data in numerical simulations. In addition, seismic wave fields, electric fields, and radar wave fields are also quite different in terms of target response sensitivity, observation methods, and geophysical field data characteristics. This results in different sensitivity, scale, and resolution of detection data to the detection target. How to achieve substantial fusion of multi-source simulation data and actual detection data characteristics is a key problem. Secondly, the training of existing pure data-driven networks relies on manual labeling, while the detection data of tunnel engineering projects are usually not exposed by excavation, and there are relatively few actual cases of comprehensive detection using multiple methods on the same face. When there is only one type of detection data or any two types of detection data, how to apply the joint inversion method of multimodal learning becomes a key issue that needs to be solved; and even if there is excavation exposure data, it is usually manually labeled, and the quality of its labels will be affected by the professional level of the labelers, making it difficult to achieve accurate labeling. Summary of the invention

[0008] In order to solve the above problems, the present invention proposes a transfer learning joint inversion method and system. The present invention constructs a stable high-quality actual data set by cleaning and gradient normalizing the actual data. For unlabeled actual detection data, a cross-domain feature alignment method based on the network end and the output end of synthetic data and actual data is proposed. The transfer learning effect between different data domains is tested by constructing a perturbation learning mode with different advanced prediction methods and training, and weakly supervised transfer learning is performed on a small amount of manually annotated actual data. The parameter learning of the network in the actual data is further optimized through model fine-tuning.

[0009] According to some embodiments, the present invention adopts the following technical solutions: A transfer learning joint inversion method comprises the following steps: Acquire detection data from electrical methods, seismic wave methods, and / or radar, and perform standardization on them, add disturbances to a portion of the detection data, and construct a data set by combining the synthetic data obtained from numerical simulations; Constructing a transfer learning network model, including a mean teacher network and a student network, wherein the student network receives synthetic data and perturbed detection data in the data set for training, and the mean teacher network receives non-perturbed detection data in the data set for training, and the prediction results of the mean teacher network are used to force the student network to produce consistent predictions on different inputs; The labeled supervised inversion loss function is used to evaluate the prediction results, and the consistency loss function is used to evaluate the difference in prediction results of the probe data in the student network and the mean teacher network; According to the actual situation on site, weak labels are used to annotate the samples revealed by excavation, and a small sample set with weak labels is constructed. The small sample set with weak labels is used to perform label training on the trained transfer learning network model and adjust the model parameters. The detection data is processed using the adjusted transfer learning network model to achieve joint inversion based on transfer learning of multi-source data and excavation exposure.

[0010] As an optional implementation, the process of obtaining detection data from electrical methods, seismic wave methods and / or radar and standardizing them includes normalizing the parameter when there is only one type of advanced detection data and assigning the same value to the other two parameters; if there are two types of detection data, taking the average of their respective normalization as the assignment of the third parameter channel.

[0011] The input parameter gradient generated by the synthetic dataset is expressed as , while the actual data based on different combinations of advance forecasting methods are ( ), is the seismic velocity gradient, is the resistivity gradient, is the dielectric constant gradient.

[0012] As an optional implementation, the parameters of the mean teacher network are updated by exponential moving average of the student network weights, specifically: ; Among them, the current teacher network parameters The teacher network parameters of the previous iteration and current student network parameters It consists of two parts. is the smoothness hyperparameter.

[0013] As an optional implementation, the transfer learning network model is forced to better utilize the probe data for training by minimizing the difference between the predictions of the student network and the predictions of the mean teacher network.

[0014] As an optional implementation, the process of evaluating the prediction results using a labeled supervised inversion loss function includes: Label supervised inversion loss function , using the mean absolute error and structural similarity index to measure the difference between the true gradient and the network predicted gradient. By comprehensively evaluating the difference, the training process of the network is guided. The calculation formula is:

[0015] - ; in, For synthetic data In the student network The prediction results, is the corresponding true label model, denote the weight parameters of mean absolute error and structural similarity respectively, represents two windows centered at (i, j, k), .

[0016] As an optional implementation, the process of using a consistency loss function to evaluate the difference in prediction results of the probe data in the student network and the mean teacher network includes: Consistency loss function , using the mean absolute error as the consistency loss to penalize the prediction differences between the teacher network and the student network on unlabeled multi-source real data. The calculation formula is: ; in, is the actual data after disturbance The prediction result obtained by inputting into the Net_stu network is The prediction result obtained by inputting the actual data into the teacher network Net_tea is: It is the feature distribution of the multimodal fusion network at the bottom layer of the decoder.

[0017] As an optional implementation, weak labels are used for the samples revealed by excavation, and the process of constructing a small sample set of weak labels includes labeling the samples based on the labeling mode of the excavation-revealed sketch.

[0018] A transfer learning joint inversion system, comprising: A data processing module is configured to obtain detection data of electrical method, seismic wave method or / and radar, and perform standardization processing on the detection data, add disturbance to part of the detection data, and construct a data set by combining the synthetic data obtained by numerical simulation; A model building and training module is configured to build a transfer learning network model, including a mean teacher network and a student network, wherein the student network receives synthetic data and perturbed detection data in the data set for training, and the mean teacher network receives non-perturbed detection data in the data set for training, and the prediction results of the mean teacher network are used to force the student network to produce consistent predictions on different inputs; The model evaluation module is configured to use a labeled supervised inversion loss function to evaluate the prediction results and a consistency loss function to evaluate the difference in prediction results of the probe data in the student network and the mean teacher network; The model fine-tuning module is configured to use weak labels to annotate the samples revealed by excavation according to the actual situation on site, construct a small sample set with weak labels, use the small sample set with weak labels to perform label training on the trained transfer learning network model, and adjust the model parameters; The joint inversion module is configured to process the detection data using the adjusted transfer learning network model to achieve transfer learning joint inversion based on multi-source data and excavation exposure.

[0019] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps in the above method are completed.

[0020] An electronic device comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps in the above method are completed.

[0021] Compared with the prior art, the present invention has the following beneficial effects: In view of the fact that in actual tunnel engineering, actual detection is affected by noise and on-site construction conditions, resulting in uneven quality of actual data, among which low-quality data will affect the network learning effect. For this reason, a set of standardized data processing processes are designed to eliminate anomalies and reduce noise in the detection data, and evaluate it from three aspects: data acquisition and processing, inversion imaging and excavation exposure comparison. Gradient features are obtained according to the standards of numerical simulation, so that the actual data can be unified with the synthetic data of numerical simulation as much as possible. A high-quality actual data set is constructed for transfer learning, so that the multi-source data joint inversion method can be better applied in actual data.

[0022] In view of the problem that there are large differences between the features of detection data and synthetic data in numerical simulation in actual tunnel engineering, the present invention proposes a model feature alignment method based on Mean-Teachers. By minimizing the difference between the measurements of the student network and the predictions of the mean teacher network, the model can be better trained with unlabeled data and improve generalization performance. In addition, by constructing a perturbation learning model with different advanced prediction methods and training to test the transfer learning effect between different data domains, the migration of the multi-source data gradient feature fusion relationship from the synthetic data domain to the actual data domain is realized.

[0023] In view of the fact that there are few comprehensive advance detection methods under the same face in actual tunnel engineering and it is difficult to manually annotate the excavation exposure situation completely and accurately on the actual data, the present invention proposes to copy the detection data, use the existing advance detection data, normalize it and assign the same value to the other two parameters, or use two existing detection data and take the average of their normalization as the value of the third parameter channel. A weakly labeled small sample transfer learning method is proposed, which reduces the difficulty of manually annotating data in tunnel engineering and reduces the cost of manual annotation through an annotation mode based on real-time photo recording or excavation exposure sketch.

[0024] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings in the specification, which constitute 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 descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0026] Figure 1 is a flow chart of the method of this embodiment; Figure 2 This is a schematic diagram of the network architecture for aligning cross-domain data feature distribution in this embodiment; Figure 3 This is a schematic diagram of the transfer learning network for fine-tuning the model in this implementation; Figure 4 It is a sketch and annotated schematic diagram of the excavation exposure of 100m in this embodiment; Figure 5 This is a comparison diagram of the actual data inversion at 100 m in this embodiment. DETAILED DESCRIPTION

[0027] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0028] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0029] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0030] In the absence of conflict, the embodiments in this application and the features in the embodiments may be combined with each other.

[0031] Embodiment 1 A joint inversion method based on transfer learning of actual multi-source detection data and excavation revelations Figure 1 As shown, the following steps are included: Step S1, based on the existing data processing methods of electrical method, seismic wave method and radar, the actual detection data is standardized to reduce interference signals, extract effective signals, and construct a high-quality actual data set.

[0032] In this embodiment, the actual data of the electrical method needs to be standardized before testing, including: According to the influence of metal bodies such as steel arches and TBM fuselages at the construction site, the extraction of effective electrical signals is achieved by adopting methods such as observation system design, metal body setting, background resistivity estimation, metal body influence correction, data adjustment and inversion parameter setting.

[0033] Before testing, actual data from earthquakes and ground-penetrating radars need to be standardized, including: The extraction of effective reflection signals is achieved by using methods such as multiple superposition, bad track processing, bandpass filtering, wave velocity estimation and direct wave removal, gain, spectrum whitening and deconvolution, FK filtering, and longitudinal and transverse wave separation. For radar dielectric constant gradient update, the DDP (Distributed Data Parallel) distributed training strategy is adopted to achieve the ability to efficiently train large models by distributing data and model parameters on multiple GPUs or multiple machines; When there is only one kind of advance detection data, the parameter is normalized and the other two parameters are assigned the same value; if there are two kinds of detection data, the average of their normalized values ​​is taken as the value of the third parameter channel.

[0034] If there are three types of detection data, just process them normally.

[0035] In this embodiment, the lack of radar detection data is used as an example to illustrate that the detection data of seismic and electrical methods are used to assign values, and the assignment method is specifically introduced. The face only uses resistivity method and seismic wave method for detection, and the detection data needs to be copied and processed. The average of the normalized seismic and electrical data is taken as the assignment of the ground penetrating radar parameter channel. In this embodiment, the conversion of resistivity data into dielectric constant requires profile extraction of the original resistivity data to make its dimensions consistent. On this basis, the grid is divided and transformed, and the 10×30 grid is converted into 100×300. To this end, this embodiment uses bilinear interpolation to expand the image, and the scaling factors in the horizontal and vertical directions are denoted as sx and sy. In this embodiment, sx and sy are both 10. For each pixel position (i, j) of the dielectric constant, it is calculated by the following formula:

[0036] Bilinear interpolation is used to calculate the value of pixel (i, j) in the dielectric constant image, and finally, the calculated pixel value is assigned to the pixel position corresponding to the dielectric constant.

[0037] This example selects a tunnel located in a weathered granite alteration area to test the multimodal learning joint inversion results based on the annotation mode of the excavation exposure sketch map. The total length of the inversion area is 100m, of which the first 30m is the public inversion area of ​​seismic and resistivity methods, and the last 70m is the non-public inversion area with only seismic inversion. Because there was no large-scale strand water during the excavation process, no resistivity detection was performed.

[0038] Step S2, construct the transfer learning network architecture, such as Figure 2 As shown, it includes a student network and a mean teacher network. The student network is the main model, which receives labeled synthetic data and unlabeled perturbed real data for training.

[0039] The mean teacher network is a moving average version of the student network during training. Its parameters are updated by the exponential moving average (EMA) of the student network weights. It mainly receives real data without disturbance. It is implemented as follows:

[0040] The teacher network parameters of the previous iteration and current student network parameters It consists of two parts, It is a smoothness coefficient hyperparameter, which usually takes a value between 0 and 1.

[0041] The predictions of the mean teacher network are used as a consistency objective to force the student network to produce consistent predictions on different inputs. By minimizing the difference between the student network's predictions and the mean teacher network's predictions, the model is forced to better utilize unlabeled data for training, thereby improving generalization performance.

[0042] Step S3, respectively calculate the prediction effects of synthetic data and actual data, input synthetic data into the student network to obtain the synthetic data prediction result, use the labeled supervised inversion loss function to evaluate the prediction result, and use the consistency loss function to evaluate the difference in prediction results of the actual data in the student network and the mean teacher network. In this example, the following labeled supervised inversion loss function and consistency loss function are used for calculation:

[0043] - ; in, For synthetic data In the student network The prediction results, is the corresponding true label model, denote the weight parameters of mean absolute error and structural similarity respectively, represents two windows centered at (i, j, k), .

[0044] ; in, is the actual data after disturbance The prediction result obtained by inputting into the Net_stu network is The prediction result obtained by inputting the actual data into the teacher network Net_tea is: It is the feature distribution of the multimodal fusion network at the bottom layer of the decoder.

[0045] Label supervised inversion loss function , which is expressed as a combination of items that comprehensively evaluate the difference between the true gradient and the network predicted gradient. The mean absolute error (MAE) and structural similarity index measure (SSIM) are used to measure the difference between the true gradient and the network predicted gradient. Consistency loss function , in order to reduce the uncertain prediction of multi-source real data in the target domain, the mean absolute error (MAE) is used as the consistency loss to penalize the prediction difference between the teacher network Net_tea and the student network Net_stu on the unlabeled multi-source real data.

[0046] Finally, the total loss function of the student network Net_stu can be defined as:

[0047] Where α and β represent the weights of labeled synthetic data and unlabeled actual data respectively, and r represents the ratio of labeled data to unlabeled data. This weight distribution takes into account the ratio of labeled data to unlabeled data. Increasing the β weight can get as close to the characteristic distribution of actual data as possible.

[0048] r is the ratio of labeled data to unlabeled data. This value will not change. β is calculated based on r. Increasing β means increasing the proportion of the unlabeled data loss function in the total loss function.

[0049] Step S4, construct a small sample set of weak labels with excavation exposure. In this embodiment, a labeling mode based on the excavation exposure sketch is adopted. Figure 5 As shown in the figure, the sketch is drawn by professionals through observation and recording of the actual excavation situation. Compared with real-time photography records, sketches can provide clearer and more accurate information because they have been carefully designed and edited. In this example, a total of 135 sets of annotated actual data were collected and drawn, of which 120 sets were used for small sample transfer learning to optimize network parameters, and 15 sets were used for testing.

[0050] This embodiment uses a weakly labeled model to adjust parameters through fine-tuning, that is, by using a small amount of weakly labeled data to train the decoder, the pre-trained model is adjusted to adapt to the distribution of actual data.

[0051] like Figure 3 As shown in the figure, in particular, the parameters of the key network layers are adjusted to make them more suitable for actual engineering data. In the fine-tuning process, the model is trained with labels using actual data, and the model parameters are adjusted to better adapt to the characteristics and distribution of actual data.

[0052] Step S5, applying the multi-source information joint inversion network that has been fine-tuned to the model to the actual data, and performing excavation verification. Figure 5 As shown in the figure, (a) and (c) are the three-dimensional visualized inversion results and two-dimensional slice diagrams of resistivity, respectively. According to the distribution of resistivity, it can be clearly seen that there is an obvious low-resistance body within 10m in front of the tunnel face, and the resistance value of the low-resistance anomaly body is quite different from the background resistivity. Combined with the water discharge of the tunnel face, it is inferred that this area is a water-rich area, filled with pressurized water and supplied by water sources. Consistent with the subsequent excavation and exposure sketch, there is a strong water-rich area in the first 10m, proving that the resistivity joint inversion result is accurate.

[0053] Figure 5 (b) and (d) are the three-dimensional visualization inversion results and two-dimensional slice diagrams of seismic wave velocity, respectively. Compared with the 30m public inversion area of ​​the electrical method, the wave velocity value of this part is 1500m / s, which is quite different from the background wave velocity. It is inferred that the rock mass in this area is broken. Figure 5 (b) The velocity distribution trend after transparency is similar to the resistivity distribution. Based on the inversion results of the common inversion area of ​​the two, it is inferred that there is an obvious broken water-bearing area in the range of 0~10m in front of the face, which is filled with pressurized water. Although the surrounding rock is broken in the range of 10~30m, it is inferred from the resistivity distribution that there is no strong water-rich area in this area. This is consistent with the information revealed by the excavation sketch and annotation. The surrounding rock in this 30m area belongs to Class V, with broken surrounding rock, developed joints and fissures, and poor stability.

[0054] The transfer learning joint inversion method based on multi-source detection data and excavation revelations used in this example has good application effects in actual engineering detection.

[0055] The main network parameters and hardware conditions in this embodiment are: the calculation is implemented on 4 graphics cards with 40G video memory. The network is built based on the PyTorch platform, the batch size of the Adam optimizer is 16, and the learning rate of the gradient optimization network parameters is set to 5e-4.

[0056] Embodiment 2 A transfer learning joint inversion system, comprising: A data processing module is configured to obtain detection data of electrical method, seismic wave method or / and radar, and perform standardization processing on the detection data, add disturbance to part of the detection data, and construct a data set by combining the synthetic data obtained by numerical simulation; A model building and training module is configured to build a transfer learning network model, including a mean teacher network and a student network, wherein the student network receives synthetic data and perturbed detection data in the data set for training, and the mean teacher network receives non-perturbed detection data in the data set for training, and the prediction results of the mean teacher network are used to force the student network to produce consistent predictions on different inputs; The model evaluation module is configured to use a labeled supervised inversion loss function to evaluate the prediction results and a consistency loss function to evaluate the difference in prediction results of the probe data in the student network and the mean teacher network; The model fine-tuning module is configured to use weak labels to annotate the samples revealed by excavation according to the actual situation on site, construct a small sample set with weak labels, use the small sample set with weak labels to perform label training on the trained transfer learning network model, and adjust the model parameters; The joint inversion module is configured to process the detection data using the adjusted transfer learning network model to achieve transfer learning joint inversion based on multi-source data and excavation exposure.

[0057] It should be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of an all-hardware embodiment, an all-software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of one or more computer-usable storage media (including but not limited to disk storage, CD - ROM , optical storage, etc.).

[0058] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0059] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0060] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0061] 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 by those skilled in the art within the spirit and principle of the present invention without creative labor shall be included in the protection scope of the present invention.

Claims

1. A transfer learning joint inversion method, characterized in that: The following steps are involved: Acquire detection data from electrical methods, seismic wave methods, and / or radar, and perform standardization on them, add disturbances to a portion of the detection data, and construct a data set by combining the synthetic data obtained from numerical simulations; Constructing a transfer learning network model, including a mean teacher network and a student network, wherein the student network receives synthetic data and perturbed detection data in the data set for training, and the mean teacher network receives non-perturbed detection data in the data set for training, and the prediction results of the mean teacher network are used to force the student network to produce consistent predictions on different inputs; The labeled supervised inversion loss function is used to evaluate the prediction results, and the consistency loss function is used to evaluate the difference in prediction results of the probe data in the student network and the mean teacher network; According to the actual situation on site, weak labels are used to annotate the samples revealed by excavation, and a small sample set with weak labels is constructed. The small sample set with weak labels is used to perform label training on the trained transfer learning network model and adjust the model parameters. The detection data is processed using the adjusted transfer learning network model to achieve joint inversion based on transfer learning of multi-source data and excavation exposure.

2. A transfer learning joint inversion method as claimed in claim 1, characterized in that: The process of obtaining the detection data of the electrical method, seismic wave method or / and radar and standardizing it includes normalizing the parameter and assigning the same value to the other two parameters when there is only one kind of advanced detection data; if there are two kinds of detection data, taking the average of their respective normalization as the assignment of the third parameter channel.

3. A transfer learning joint inversion method as claimed in claim 1, characterized in that: The input parameter gradient generated by the synthetic dataset is expressed as , while the actual data based on different combinations of advance forecasting methods are ( ), is the seismic velocity gradient, is the resistivity gradient, is the dielectric constant gradient.

4. A transfer learning joint inversion method as claimed in claim 1, characterized in that: The parameters of the mean teacher network are updated by the exponential moving average of the student network weights, specifically: ； Among them, the current teacher network parameters The teacher network parameters of the previous iteration and current student network parameters It consists of two parts. is the smoothness hyperparameter.

5. The transfer learning joint inversion method according to claim 1, characterized in that: By minimizing the difference between the student network’s predictions and the mean teacher network’s predictions, the transfer learning network model is forced to better utilize the probe data for training.

6. A transfer learning joint inversion method as claimed in claim 1, characterized in that: The process of evaluating the prediction results using the labeled supervised inversion loss function includes: Label supervised inversion loss function , using the mean absolute error and structural similarity index to measure the difference between the true gradient and the network predicted gradient. By comprehensively evaluating the difference, the training process of the network is guided. The calculation formula is: - ； in, For synthetic data In the student network The prediction results, is the corresponding true label model, denote the weight parameters of mean absolute error and structural similarity respectively, represents two windows centered at (i, j, k), .

7. A transfer learning joint inversion method as claimed in claim 1, characterized in that: The process of using the consistency loss function to evaluate the difference in the prediction results of the probe data in the student network and the mean teacher network includes: Consistency loss function , using the mean absolute error as the consistency loss to penalize the prediction differences between the teacher network and the student network on unlabeled multi-source real data. The calculation formula is: ； in, is the actual data after disturbance The prediction result obtained by inputting into the Net_stu network is The prediction result obtained by inputting the actual data into the teacher network Net_tea is: It is the feature distribution of the multimodal fusion network at the bottom layer of the decoder.

8. A transfer learning joint inversion method as claimed in claim 1, characterized in that: Weak labels are used to annotate the samples revealed by excavation, and the process of constructing a small sample set with weak labels includes annotating samples based on the annotation mode of the excavation-revealed sketch.

9. A transfer learning joint inversion system, characterized in that: include: A data processing module is configured to obtain detection data of electrical method, seismic wave method or / and radar, and perform standardization processing on the detection data, add disturbance to part of the detection data, and construct a data set by combining the synthetic data obtained by numerical simulation; A model building and training module is configured to build a transfer learning network model, including a mean teacher network and a student network, wherein the student network receives synthetic data and perturbed detection data in the data set for training, and the mean teacher network receives non-perturbed detection data in the data set for training, and the prediction results of the mean teacher network are used to force the student network to produce consistent predictions on different inputs; The model evaluation module is configured to use a labeled supervised inversion loss function to evaluate the prediction results and a consistency loss function to evaluate the difference in prediction results of the probe data in the student network and the mean teacher network; The model fine-tuning module is configured to use weak labels to annotate the samples revealed by excavation according to the actual situation on site, construct a small sample set with weak labels, use the small sample set with weak labels to perform label training on the trained transfer learning network model, and adjust the model parameters; The joint inversion module is configured to process the detection data using the adjusted transfer learning network model to achieve transfer learning joint inversion based on multi-source data and excavation exposure.

10. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the steps of the method according to any one of claims 1 to 8 are completed.

Images