Bi-LSTM network dispersion curve inversion method, medium and equipment

Abstract

The present invention relates to a dispersion curve inversion method, medium and equipment based on a Bi-LSTM network, and belongs to the field of geophysical exploration data processing technology. The method first constructs a Bi-LSTM network, which splices two independent conventional LSTM networks in different directions together as a main network unit, and splices the outputs of the two networks as the final output, so that the network can simultaneously obtain the previous information of the past moment and the following information of the future moment, so that it has a better effect when processing sequence data. In addition, the present invention introduces physical constraints in the training process, firstly takes the real dispersion curve as input, predicts the shear wave velocity model through the network, and then uses the shear wave velocity model to perform the dispersion curve, and introduces the difference between the two dispersion curves into the loss function in a weighted manner as a physical constraint, thereby improving the prediction accuracy and stability of the network.

Description

Technical Field

[0001] The present invention belongs to the field of exploration geophysics, and in particular relates to a Bi-LSTM network dispersion curve inversion method, medium and equipment. Background Art

[0002] Near-surface exploration is an important research direction in geophysical exploration. Its purpose is to obtain the structural characteristics of shallow strata and physical parameters such as longitudinal and shear wave velocities. Compared with the deep geophysical exploration methods used in the field of oil and gas exploration, near-surface exploration is mainly used in the fields of low-velocity reduction zone determination and engineering geophysical exploration. It is currently the most direct and effective way to solve engineering construction problems. Using surface wave exploration to obtain the shallow surface shear wave velocity structure is a major research direction.

[0003] At present, the surface wave multi-channel analysis method is widely used in the investigation of shallow surface shear wave velocity structure. Its basic workflow is to transform the record to obtain the dispersion energy map, pick up the dispersion curve, and finally invert the dispersion curve to obtain the velocity model. There are many inversion optimization algorithms for dispersion curves, mainly including least squares method, genetic algorithm, fast simulated annealing method, hot bath simulated annealing method, particle swarm algorithm, hybrid global optimization algorithm, cuckoo search algorithm, wave equation dispersion curve inversion algorithm based on steepest descent method, and preconditioned truncated Newton method. However, the above methods often rely on the initial model and prior conditions, and the calculation speed is often low.

[0004] With the development of artificial intelligence, deep learning methods have played a great role in many fields such as natural language understanding and visual target recognition. More and more scholars have begun to use deep learning methods to solve geophysical problems. In recent years, experts and scholars at home and abroad have gradually applied deep learning methods to dispersion curve inversion. At present, the existing neural network for intelligent inversion of dispersion curves is mainly a conventional LSTM network, but this network can only obtain past information. However, for actual sequence data, the information at a certain moment is often not only related to the information at the past moment, but also to the information at the future moment. That is, when processing data, not only the past information is often needed, but also the future information. Therefore, a network that can obtain both the past moment and the future moment is needed.

[0005] In addition, existing deep learning-based intelligent inversion techniques for dispersion curves often directly use the loss between the predicted velocity model and the true velocity model for backpropagation and updating network parameters. They lack physical constraints during the training process and are strongly dependent on the true velocity model, which results in poor generalization of the network and low prediction accuracy. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a new dispersion curve inversion method based on Bi-LSTM network.

[0007] The present invention is implemented in this way: a method for inverting a dispersion curve based on a Bi-LSTM network, the method comprising the following steps:

[0008] The first step is to build a Bi-LSTM network model. Two independent conventional LSTM networks with different directions are spliced ​​together as the main network unit, and the outputs of the two networks are spliced ​​as the final output. The LSTM unit in the network consists of a forget gate, an input gate, an output gate and a cell state.

[0009] This design allows the network to simultaneously obtain previous information from the past and next information from the future, which can not only solve the problem of poor long-term dependency of traditional recurrent neural networks, but also better extract features from sequence data;

[0010] The second step is to design a shear wave velocity model and obtain the fundamental order dispersion curve of Rayleigh surface waves through forward modeling. The fundamental order dispersion curve of Rayleigh surface waves is used as the input of the data set, so as to construct a feature-rich data set for the next step of network model training.

[0011] Furthermore, the shear wave velocity model includes an increasing velocity model, a velocity model containing a high-speed hard interlayer, or a velocity model containing a low-speed soft interlayer.

[0012] The third step is to use the corresponding equipment to obtain exploration data, pre-process the data set, and divide the data set into training set and test set in a ratio of 20:1;

[0013] Step 4: Use the training set to train the network model constructed in the first step; first use the traditional mean square loss function to train to obtain a pre-trained network model, and then use the improved mean square loss function to continue training based on the pre-trained network model to fine-tune the network model; the improved mean square loss function It can be expressed as:

[0014] ;

[0015] ;

[0016] is the true shear wave velocity model, To predict the shear wave velocity model, N is the number of sample pairs in a batch, is the input fundamental order dispersion curve, is the fundamental dispersion curve being performed using the predicted shear wave velocity model, Represents weight.

[0017] Furthermore, the traditional mean square loss function It can be expressed as:

[0018] ;

[0019] in, is the true shear wave velocity model, To predict the shear wave velocity model, N is the number of sample pairs in a batch.

[0020] Step 5: Use the simulated test set and actual data to test or process the trained network model.

[0021] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the Bi-LSTM network dispersion curve inversion method.

[0022] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the Bi-LSTM network dispersion curve inversion method.

[0023] The beneficial effects of the present invention compared with the prior art are as follows:

[0024] The present invention develops a Bi-LSTM network dispersion curve inversion method, which can realize the inversion of Rayleigh surface wave fundamental order dispersion curve with high efficiency and high precision;

[0025] The present invention innovatively applies the Bi-LSTM network to the intelligent inversion process of dispersion curves. This network splices two independent conventional LSTM networks in different directions together, and splices the outputs of the two networks as the final output. The LSTM unit in the network consists of a forget gate, an input gate, an output gate and a cell state. Through this design, the network can simultaneously obtain the context information of the past moment and the context information of the future moment, which can not only solve the problem of poor long-term dependence of traditional recurrent neural networks, but also better extract features from sequence data.

[0026] In addition, the present invention introduces physical constraints during the training process, takes the real dispersion curve as input, predicts the shear wave velocity model through the network, and then uses the shear wave velocity model to perform the dispersion curve, obtains the error between it and the input dispersion curve, and introduces it into the loss function in the form of weights as physical constraints, thereby improving the prediction accuracy and stability of the network. Through this innovation, the present invention overcomes the deficiency of the traditional intelligent dispersion curve inversion method that lacks physical constraints, thereby improving the interpretability of the training process and the prediction accuracy of the network. At the same time, this innovation also enables the present invention to get rid of the shortcomings of the traditional dispersion curve intelligent inversion method that is heavily dependent on the initial velocity model and has low efficiency.

[0027] The present invention has good inversion effects on shallow surface velocity models in three cases: velocity increasing model, velocity model containing high-speed hard interlayer and velocity model containing low-speed soft interlayer. It can realize high-precision dispersion curve inversion and perform efficient processing of surface wave data, thereby providing effective guidance and assistance for shallow surface exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is the workflow diagram of intelligent inversion of dispersion curve;

[0029] Figure 2 This is a schematic diagram of the structure of the Bi-LSTM network;

[0030] Figure 3 It is a schematic diagram of the structure of the LSTM unit;

[0031] Figure 4 These are 10 randomly selected training sample images, where (a) is the shear wave velocity model and (b) is the dispersion curve;

[0032] Figure 5 is a loss curve graph, wherein (a) is the loss curve before 1000 training times (i.e., using the conventional loss function), (b) is the loss curve when the training times are between 500 and 1000 times, and (c) is the loss curve when the training times are between 1000 and 2000 times (i.e., using the improved loss function of the present invention);

[0033] Figure 6: are test result diagrams of the simulation test set, where (a) is a comparison between the predicted velocity model and the true shear wave velocity model of sample 1, where the velocity increases nearly linearly with depth; (b) is a comparison between the fundamental order dispersion curve obtained by forward modeling the predicted velocity model of sample 1 and the true dispersion curve; (c) is a comparison between the predicted velocity model and the true shear wave velocity model of sample 2 with a significant high-speed layer; (d) is a comparison between the fundamental order dispersion curve and the true dispersion curve obtained by forward modeling the predicted velocity model of sample 2 with a significant low-speed layer; (e) is a comparison between the predicted velocity model and the true shear wave velocity model of sample 3 with a significant low-speed layer; (f) is a comparison between the fundamental order dispersion curve and the true dispersion curve obtained by forward modeling the predicted velocity model of sample 3;

[0034] Figure 7 It is a single shot seismic record diagram of some actual data;

[0035] Figure 8 This is the inversion result diagram of the dispersion curve corresponding to the actual single-shot seismic record;

[0036] Fig. 9 It is a pseudo two-dimensional velocity profile obtained by inverting actual data using the method of the present invention;

[0037] Fig.10 It is a pseudo-two-dimensional velocity profile obtained by inverting actual data using traditional intelligent inversion methods without physical constraints;

[0038] Fig.11 It is a pseudo two-dimensional velocity profile obtained by inverting the actual data using the damped least squares method. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples. It should be understood that the specific examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] The present invention provides a new dispersion curve inversion technology based on Bi-LSTM network, which successfully realizes high-precision shallow surface velocity inversion. The basic idea of ​​the present invention is to take the real dispersion curve as input, predict the shear wave velocity model through the network, and then use the shear wave velocity model to perform the dispersion curve, find its error with the input dispersion curve, and introduce it into the loss function in the form of weights as a physical constraint, thereby improving the prediction accuracy and stability of the network. The network used in the present invention is a Bi-LSTM network, which splices two independent LSTM networks in different directions together, and splices the outputs of the two networks as the final output; through this design, the network can simultaneously obtain the previous information of the past moment and the following information of the future moment, so that it has a better effect when processing sequence data. The present invention realizes high-precision inversion of simulated data and actual data, and the present invention is described in detail below in conjunction with the accompanying drawings.

[0041] like Figure 1 As shown, the Bi-LSTM network dispersion curve inversion method proposed in the present invention has the following specific steps:

[0042] Step S10: Construct a Bi-LSTM network model, splice two independent conventional LSTM networks in different directions together, and splice the outputs of the two networks as the final output. The network structure is as follows: Figure 2 As shown. Through this design, the network can simultaneously obtain the previous information of the past moment and the following information of the future moment, so that it has a better effect in processing sequence data. The LSTM unit in the network consists of a forget gate, an input gate, an output gate and a cell state. Its specific structure is as follows Figure 3 As shown in Figure 2. Through this design, the network can not only solve the problem of poor long-term dependency of traditional recurrent neural networks, but also better extract features from sequence data.

[0043] Step S20: Design a shear wave velocity model and forward model the fundamental order dispersion curve of Rayleigh surface waves. In order to make the Bi-LSTM network applicable to various velocity models and have stronger generalization ability, the present invention designs various types of shear wave velocity models, mainly considering velocity models in three cases, namely, an increasing velocity model, a velocity model with a high-speed hard interlayer, and a velocity model with a low-speed soft interlayer. 4200 velocity models are randomly generated as labels, and the velocity models are forward modeled to obtain the fundamental order dispersion curve of Rayleigh surface waves as the input of the data set, thereby constructing a feature-rich data set for the next step of network model training. Figure 4 are 10 training samples randomly selected from the training set, among which, Figure 4 (a) is the shear wave velocity model, i.e. the label of the data set. Figure 4(b) in the figure is the dispersion curve, which is the input of the data set.

[0044] Step S30: normalize and resample the earthquake data being performed by the above model, and divide it into a training set and a test set at a ratio of 20:1 to complete the data set production.

[0045] Step S40: Send the training set to the above Bi-LSTM network for training. The maximum number of training epochs is 2000, the batch size is 32, and the initial learning rate is 0.001. In the present invention, the conventional mean square loss function is first used to train 1000 times to obtain a pre-trained network model, and then the improved mean square loss function proposed in the present invention is used to continue training 1000 times based on the pre-trained network model to fine-tune the network model and improve the prediction accuracy of the network. For one batch, the traditional mean square loss function It can be expressed as:

[0046] ;

[0047] in, is the true shear wave velocity model, To predict the shear wave velocity model, N is the number of sample pairs in a batch. On this basis, the present invention performs a dispersion curve again for the velocity model predicted by the network model during the training process, and calculates the mean square loss between the dispersion curve of the predicted velocity model and the input dispersion curve as the weight :

[0048] ;

[0049] in, is the input fundamental order dispersion curve, is the fundamental dispersion curve being performed using the predicted shear wave velocity model. After calculating the weights, the weights are added to the traditional mean square loss function to obtain the improved mean square loss function , thereby improving the stability of the network and the prediction accuracy:

[0050] ;

[0051] During the training process, the Adam (Adaptive momentum) gradient descent method is used for training. This method can design independent adaptive learning rates for different parameters by calculating the first-order moment estimate of the gradient (the average of all past gradients and the current gradient) and the second-order moment estimate (the average of the square of the past gradient and the square of the current gradient), making it easier for the parameters to converge to the optimal solution in the later stages of training.

[0052] The loss curve before 1000 training times (i.e. using the conventional loss function) is as follows Figure 5 As shown in (a) in the figure, we can see that the loss value gradually decreases with the increase of training times, and finally gradually stabilizes. The loss function for training times between 500 and 1000 is as follows: Figure 5 As shown in (b) in the figure, it can be seen that the loss value no longer has a downward trend after about 650 times, but fluctuates continuously within a small range, which means that the network performance has stabilized at this time and no longer needs to be optimized. The loss curve for training times of 1000-2000 times (i.e., using the improved loss function of the present invention) is shown in Figure 5 As shown in (c) in the figure, it can be seen that after using the improved loss function, the loss curve begins to decline further, which means that the performance of the network has been further optimized. This is because the improved loss function of the present invention introduces the dispersion curve error as a weight into the loss function, so that the loss weight of samples with poor training effect (large dispersion curve error) is amplified and highlighted, so that the network model can be further fine-tuned and the network performance can be further optimized.

[0053] In addition, in order to compare the effects of the traditional intelligent inversion method without physical constraints and the method proposed in the present invention, the present invention also used the conventional mean square loss function without physical constraints to obtain the network model with the same training set, the same hyperparameters, and the same optimizer for 2000 training times, so as to compare the effects with the method proposed in the present invention.

[0054] The training environment for this example is:

[0055] (1) Processor: 13th Gen Intel(R) Core(TM) i7-13700KF;

[0056] (2) Memory: 64 GB;

[0057] (3) System type: 64-bit operating system;

[0058] (4) Graphics card: NVIDIA GeForce RTX4090, 24GB;

[0059] (5) Platform: PyTorch neural network framework based on PyCharm compiler.

[0060] Step S50: Use the simulated test set and actual data to test the trained network model

[0061] Figure 6 The test results of the method of the present invention using a simulated test set are shown here, only 3 samples are shown. Figure 6(a) is a comparison between the predicted velocity model and the true shear wave velocity model of sample 1, where the velocity increases nearly linearly with depth (the solid line is the true velocity model, and the dotted line is the predicted velocity model); Figure 6 (b) is the comparison between the fundamental dispersion curve obtained by forward modeling the predicted velocity model of sample 1 and the true dispersion curve; Figure 6 (c) is the comparison between the predicted velocity model and the true shear wave velocity model of sample 2 with a significant high-speed layer; Figure 6 (d) is the comparison between the fundamental dispersion curve obtained by forward modeling the predicted velocity model of sample 2 and the true dispersion curve; Figure 6 (e) is the comparison between the predicted velocity model and the true shear wave velocity model of sample 3 with a significant low-velocity layer; Figure 6 (f) is a comparison between the fundamental dispersion curve obtained by forward modeling the predicted velocity model of sample 3 and the real dispersion curve. By observing the test results of the simulation data, it can be seen that the method proposed in the present invention has a high inversion accuracy.

[0062] The trained network model was tested using actual data from the underground soil cave detection area in Xiaokao Village, Liqianhu Town, Tieling County. The actual data was collected in a rolling arrangement, with a total of 14 rolls, each roll of 4m, a total of 15 arrangements, and the observation system of each arrangement was the same, with a minimum offset of 4m, a track spacing of 1m, 12 tracks, 2048 sampling points, and a sampling interval of 0.2ms. Among them, the actual seismic record of a certain arrangement is as follows Figure 7 As shown in the figure, the dispersion curve of this earthquake record is inverted, and the predicted shear wave velocity model is obtained as follows Figure 8 As shown in the figure, observing the prediction results, we can see that there is an obvious low-velocity layer in the dotted box part of the figure. Combined with the previous investigation results, it is found that it corresponds to the underground soil cave in the work area.

[0063] The method of the present invention is used to invert the 15 arrangements, and the inversion results are interpolated into pseudo two-dimensional profiles, such as Fig. 9 The pseudo 2D profile predicted by the network model trained using only the conventional mean square loss function without physical constraints is shown in Fig.10 As shown in the figure, the pseudo two-dimensional profile obtained using the damped least squares method is as follows Fig.11 As shown. By comparing the three pictures, we can find Fig. 9 Compared to Fig.10 The abnormal body can be distinguished more clearly, and its shallow interface morphology is compared with Fig.10 Closer Fig.11This shows that the network model trained by the training method combined with the improved mean square loss function proposed in the present invention can more accurately depict the underground velocity structure, that is, has a higher resolution, compared with the network model trained by the traditional training method using only the conventional mean square loss function. Therefore, it is believed that the method of the present invention has higher prediction accuracy and practical promotion and application value.

[0064] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A dispersion curve inversion method based on Bi-LSTM network, characterized in that: The method comprises the following steps: The first step is to build a Bi-LSTM network model. Two independent conventional LSTM networks with different directions are spliced ​​together as the main network unit, and the outputs of the two networks are spliced ​​as the final output. The LSTM unit in the network consists of a forget gate, an input gate, an output gate and a cell state. The second step is to design the shear wave velocity model and obtain the fundamental order dispersion curve of Rayleigh surface wave by forward modeling, and obtain the fundamental order dispersion curve of Rayleigh surface wave as the input of the data set; The third step is to use the corresponding equipment to obtain exploration data, pre-process the data set, and divide the data set into training set and test set in a ratio of 20:1; Step 4: Use the training set to train the network model constructed in the first step; first use the traditional mean square loss function to train to obtain a pre-trained network model, and then use the improved mean square loss function to continue training based on the pre-trained network model to fine-tune the network model; the improved mean square loss function L b * It is expressed as: y i is the true shear wave velocity model, To predict the shear wave velocity model, N is the number of sample pairs in a batch, z i is the input fundamental order dispersion curve, is the fundamental dispersion curve being performed using the predicted shear wave velocity model, w i represents weight; Step 5: Use the simulated test set and actual data to test or process the trained network model.

2. According to claim 1, a dispersion curve inversion method based on Bi-LSTM network is characterized in that: The shear wave velocity model includes an increasing velocity model and a velocity model containing a high-speed hard interlayer, or the shear wave velocity model includes an increasing velocity model and a velocity model containing a low-speed soft interlayer.

3. According to claim 1, a method for inverting dispersion curves based on a Bi-LSTM network is characterized in that: The traditional mean square loss function L b It is expressed as: Among them, y i is the true shear wave velocity model, To predict the shear wave velocity model, N is the number of sample pairs in a batch.

4. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the Bi-LSTM network dispersion curve inversion method according to any one of claims 1 to 3.

5. A computer device, characterized in that: The device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the Bi-LSTM network dispersion curve inversion method according to any one of claims 1 to 3.

Images