RBF Inversion Method under Special Reservoir Sequence Constraints

By employing the RBF inversion method under special reservoir sequence constraints, and utilizing well logging and seismic data for hierarchical constraints and adaptive optimization, the problem of insufficient resolution in thin reservoirs is solved, and more accurate low-frequency models and oil and gas reservoir predictions are achieved.

CN118068426BActive Publication Date: 2026-11-17CHINA NAT PETROLEUM CORP +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211412689.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-11-11
Publication Date
2026-11-17
Estimated Expiration
2042-11-11

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively distinguish the vertical and lateral variations of thin reservoirs in unconventional oil and gas reservoirs. Seismic data resolution is insufficient, and low-frequency model building methods suffer from bullseye effect and multiple solutions. Existing RBF networks have weak classification capabilities when there is significant inter-class overlap and large intra-class distance, and they cannot adapt to changes in the external environment.

Method used

A special reservoir sequence constraint-based RBF inversion method is adopted. By preprocessing well logging curves and focusing seismic data on faults, the volcanic eruption cycles and sequence boundaries are constrained step by step. Combined with the RBF neural network learning algorithm, the number of RBFs and the center vector are adaptively adjusted, and the low-frequency model is optimized by gradient descent training method.

Benefits of technology

A low-frequency model that more closely approximates the actual underground changes has been established, improving the accuracy and reliability of reservoir prediction and enabling comprehensive seismic and geological prediction of unconventional oil and gas, thus adapting to changes in the external environment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118068426B_ABST
    Figure CN118068426B_ABST
Patent Text Reader

Abstract

This invention discloses a RBF inversion method under special reservoir sequence constraints, belonging to the field of oil and gas seismic exploration technology. The method includes: S1, well logging curve preprocessing: performing well logging curve environmental correction, multi-well consistency processing, and multi-well standardization on the original well logging curves; simultaneously performing fault focusing processing on seismic data to obtain seismic data with improved accuracy; using the processed well logging curves and seismic data to perform fine calibration of wells and seismic data; S2, delineating volcanic rock sequence boundaries and constructing a stratigraphic framework; S3, establishing an initial model based on well data and seismic data using the RBF algorithm; S4, using a multi-attribute fused data volume as input, establishing a low-frequency model under volcanic rock sequence constraints based on the obtained initial model, performing phase-controlled inversion, and predicting high-quality volcanic rock reservoirs. The low-frequency model ultimately obtained by this invention more closely approximates the actual underground variation patterns and can achieve comprehensive seismic and geological prediction of unconventional oil and gas.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to an RBF inversion method under special reservoir sequence constraints, belonging to the field of oil and gas seismic exploration technology. Background Technology

[0002] Currently, oil and gas exploration has gradually shifted from conventional oil and gas exploration to exploration of unconventional and special reservoirs. Unconventional and special reservoirs are influenced vertically by variations in eruption periods, resulting in thin reservoirs with multiple stacked layers. Laterally, they are affected by sedimentary facies zones and lithological variations, leading to rapid reservoir changes. Therefore, the requirements for reservoir prediction accuracy are increasingly stringent. Due to limitations in the dominant frequency of seismic data, seismic resolution and post-stack impedance inversion cannot effectively distinguish these thin reservoirs.

[0003] Radial Basis Function (RBF) neural networks are widely used in function approximation, system identification, and pattern recognition due to their simple structure and fast learning speed. The key to constructing an RBF network is the appropriate selection of the number of radial basis functions and their center vectors. Commonly used methods in existing technologies include K-means clustering and C-Means algorithm. These methods manually determine the number of radial basis functions and initial vectors, then use the nearest neighbor method based on Euclidean distance to achieve clustering. However, when faced with situations involving large inter-class overlap and large intra-class distances, the classification ability of these methods is severely weakened. Furthermore, the training and operation processes of the network are completely independent; the system's characteristics change with the external environment, requiring retraining whenever changes occur. Therefore, the methods used in existing technologies are not conducive to the widespread application of these networks.

[0004] For reservoir inversion, low-frequency models can enhance the spatial identification capability of geological bodies and improve the accuracy and reliability of absolute impedance inversion. There are three main conventional methods for establishing low-frequency models: one relies on well-based mathematical interpolation algorithms. In areas with few wells and large well spacing, this method is prone to bullseye phenomena. Furthermore, when recursively extrapolating low-frequency components along the seismic profile, the low-frequency information is simply interpolated or extrapolated linearly, unaffected by structural variations such as stratigraphic attitude and lithological changes. For areas with complex structures, multiple vertical stackings, and significant lateral variations in lithology and physical properties, simple interpolation and extrapolation cannot reflect the lateral variations of complex geological bodies, leading to inaccurate inversion results. Another method uses velocity volume constraints to construct low-frequency models. This method can only compensate for extremely low-frequency information (less than 2Hz) and cannot compensate for the seismic frequency band information (2-10Hz) that is completely missing from conventional seismic data. A third method uses iterative methods to construct low-frequency models, but this method is prone to causing multiple solutions in the low-frequency model.

[0005] In summary, there is an urgent need for a technical method that can overcome the above-mentioned shortcomings, take into account the influence of multiple factors such as eruption period, cycle, lithofacies, and paleogeography, establish a low-frequency model that is closer to the actual underground change law, and realize the comprehensive seismic and geological prediction of unconventional oil and gas. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention aims to provide an RBF inversion method under special reservoir sequence constraints, in order to obtain a more accurate and reliable low-frequency model and achieve the goal of comprehensive seismic and geological prediction of unconventional oil and gas.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A special reservoir sequence constraint RBF inversion method, comprising the following steps performed sequentially:

[0009] S1. Well logging curve preprocessing: Perform well logging curve environmental correction, multi-well consistency processing and multi-well standardization processing on the original well logging curves to obtain preprocessed well logging curves and well data, determine the target layer in the study area, and statistically analyze the VSP formation velocity.

[0010] The VSP refers to the vertical seismic profile.

[0011] Fault focusing processing: interpretive preprocessing of seismic data, filtering out high-frequency components that are not fault features at fault locations to obtain seismic data and seismic data with improved accuracy;

[0012] Fine calibration of well and seismic data was performed using pre-processed logging curves and seismic data.

[0013] S2. Based on the preprocessed well logging curves and preprocessed seismic data, volcanic eruption cycles and volcanic eruption sequence boundaries are constrained step by step. At the same time, based on the fine calibration of wells and seismic data, volcanic rock sedimentary discontinuities are identified, volcanic rock sequence boundaries are delineated, and a stratigraphic framework is constructed.

[0014] S3. Based on the established stratigraphic framework, an initial model is built using the RBF algorithm based on well data and seismic data;

[0015] S4. Using the multi-attribute fusion data volume as input, a low-frequency model is established under the constraint of volcanic rock sequence based on the obtained initial model, and phase control inversion is carried out to predict high-quality volcanic rock reservoirs.

[0016] As a limitation on step S3: step S3 includes the following steps performed sequentially:

[0017] a1) Apply the RBF neural network learning algorithm to preprocess well data and seismic data respectively, and record the output results;

[0018] a2) Based on the output results obtained in step a1), calculate the output error generated by each input vector, and combine the distribution of the error in the input space with the effect of each RBF on the network to perform adaptive RBF optimization prediction.

[0019] a3) Based on the adaptive RBF optimization prediction results obtained in step a2), the results are compared with well-seismic data, and the joint loss function is calculated using the gradient descent training method. If the joint loss function does not converge, the process is restarted from step a2) until the results converge, and the initial model can be output.

[0020] As a further constraint on step S3, the joint loss function in step a3) is defined as:

[0021]

[0022] In the formula, x n Input values ​​for the sample points, t n This represents the output of the corresponding sample in layer k, where N is the number of sample points, N0 is the number of hidden unit layers, and N... kH Let ω be the number of hidden units in the k-th layer. jk φ is the output weight of sample j in the k-th layer. j (x n ) is an implicit unit function, and n, i, j, k are all index subscript variables.

[0023] The relative width of W and its weight ω ji Negative gradient:

[0024]

[0025] Where c i σ is the center of the i-th Gaussian component. i Let L be the radius of the i-th Gaussian component, and L and U be the space where the sample is located.

[0026] The adjustment methods for the width and weights of the radial basis function along the gradient descent direction are as follows:

[0027]

[0028] In the formula, η and β∈[0,1] are the learning efficiencies of the width and weight, respectively, which determine the gradient search step size; t is the number of iterations.

[0029] By adopting the above-described technical solution, the beneficial effects achieved by this invention compared to the prior art are as follows:

[0030] (1) The low-frequency model obtained by the RBF neural network learning algorithm in this invention is closer to the real underground change law and has geological significance. It provides an accurate and reliable low-frequency model for subsequent reservoir prediction and realizes the comprehensive seismic and geological prediction of unconventional oil and gas.

[0031] (2) The present invention applies the RBF neural network learning algorithm. It does not require the number of RBFs and the center vector to be determined in advance. The number of RBFs can be adaptively adjusted according to the distribution of error in the input space, and the center vector can be adjusted appropriately so that the approximation performance and generalization ability of the network can meet the requirements of high quality, thus optimizing the model. At the same time, the training and operation of the network can be carried out alternately, so that it can adapt to the slow changes of the external environment.

[0032] (3) The gradient descent training method is used in this invention to minimize the error function at the fastest speed, complete the mapping from input to output, thereby training the width and weight of the hidden layer, while avoiding the problem of local optimization.

[0033] In summary, this invention makes full use of drilling, core, and seismic data, and from a new perspective, by fully considering the vertical variation of special reservoirs due to eruption periods and the horizontal influence of sedimentary facies zones and lithological changes, and by optimizing through step-by-step constraints, establishes a low-frequency model that is closer to the actual underground variation patterns, thereby achieving comprehensive seismic and geological prediction of unconventional oil and gas. Attached Figure Description

[0034] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0035] Figure 1 This is a flowchart of an RBF inversion method under hierarchical constraints according to an embodiment of the present invention;

[0036] Figure 2 This is a flowchart of RBF modeling according to an embodiment of the present invention;

[0037] Figure 3 (a) Impedance plane diagram of the target layer in the low-frequency model established for conventional well interpolation;

[0038] Figure 3 (b) is the impedance plane diagram of the target segment in RBF modeling under the hierarchical constraints of this embodiment of the invention;

[0039] Figure 4 (a) is the in-well interpolation impedance inversion diagram;

[0040] Figure 4 (b) is the RBF inversion diagram under the hierarchical constraints of an embodiment of the present invention. Detailed Implementation

[0041] To better explain and facilitate understanding of the present invention, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0042] Example: RBF Inversion Method under Special Reservoir Sequence Constraints

[0043] This embodiment provides a special reservoir sequence constraint RBF inversion method, which has been applied to the prediction of unconventional reservoirs in the Junggar Basin.

[0044] like Figure 1 As shown, the RBF inversion method under special reservoir sequence constraints includes the following steps performed sequentially:

[0045] S1. Well logging curve preprocessing: Perform well logging curve environmental correction, multi-well consistency processing and multi-well standardization processing on the original well logging curves to obtain preprocessed well logging curves and well data, determine the target layer in the study area, and statistically analyze the VSP formation velocity.

[0046] The VSP refers to the vertical seismic profile.

[0047] Fault focusing processing: interpretive preprocessing of seismic data, filtering out high-frequency components that are not fault features at fault locations to obtain seismic data and seismic data with improved accuracy;

[0048] Fine calibration of well and seismic data was performed using pre-processed logging curves and seismic data.

[0049] S2. Based on the preprocessed well logging curves and preprocessed seismic data, volcanic eruption cycles and volcanic eruption sequence boundaries are constrained step by step. At the same time, based on the fine calibration of wells and seismic data, volcanic rock sedimentary discontinuities are identified, volcanic rock sequence boundaries are delineated, and a stratigraphic framework is constructed.

[0050] S3. Based on the established stratigraphic framework, an initial model is built using the RBF algorithm based on well data and seismic data;

[0051] S4. Using the multi-attribute fusion data volume as input, a low-frequency model is established under the constraint of volcanic rock sequence based on the obtained initial model, and phase control inversion is carried out to predict high-quality volcanic rock reservoirs.

[0052] In predicting unconventional reservoirs in the Junggar Basin, preprocessed well and seismic data were used, with weathering crust and sedimentary interlayers as marker layers. A well-to-well sequence stratigraphy was conducted, dividing the volcanic eruptions into three phases. The basal overflow facies of the Bashan Formation (C2b) was used as the regional caprock, and the Songkarsu Formation b section (C1sb) consisted of lava and tuff as interlayers. The eruption phases were established using a combination of well data, identifying a total of three eruptions, which then guided the development of a low-frequency Carboniferous model.

[0053] In this embodiment, since the same strata in the same block have similar depositional environments, depositional ages, and stratigraphic properties, their distribution patterns on the histogram are similar. Therefore, the original logging curves are adjusted accordingly. Specifically, by statistically analyzing the distribution of the histogram of the target layer curves, the deviation values ​​of each well are calculated, corrected, and high-quality curves and well data suitable for subsequent reservoir inversion are obtained. The preprocessed logging curves lay the foundation for subsequent sequence stratigraphy, fine well-seismic calibration, and low-frequency model establishment.

[0054] Since volcanic activity and the development of volcanic structures are usually closely related to regional fault activity, and favorable facies zones of volcanic rocks are usually distributed along fault zones, fault identification of seismic profiles is crucial. Fault characteristics in seismic profiles are usually manifested as the faulting of the same phase axis, which is mostly due to the low-frequency characteristics of earthquakes. Therefore, fault focusing processing of seismic data is necessary.

[0055] In practice, interpretive preprocessing is performed on the seismic data. High-frequency components that do not exhibit fault characteristics are filtered out at fault locations, resulting in improved seismic data. In this embodiment, high-frequency components that do not exhibit fault characteristics are filtered out at fault locations in the seismic profile, while relatively low-frequency components are retained to better highlight the fault morphology. Here, "relatively low frequency" refers to frequencies lower than the high frequencies of non-fault characteristics. The processing mainly involves setting key parameters such as low frequency, high frequency, and focusing frequency. The focusing frequency must be between the low and high frequencies, while the high frequency can be appropriately lower than the dominant seismic frequency. Here, the low frequency is less than 8 Hz, and the high frequency is greater than 30 Hz. Fine-tuned well-seismic calibration is then performed using the preprocessed well logging curves and seismic data.

[0056] Based on the preprocessed well logging curves from step S1 and the improved seismic data, volcanic eruption cycles and sequence boundaries are constrained at each level. Volcanic sequence stratigraphy is mainly divided into three levels: volcanic eruption cycles (Level I sequence), volcanic eruption phases (Level II sequence), and volcanic eruption rhythms (Level III sequence).

[0057] Among them, volcanic eruption cycles are controlled by volcanic rock formations, and sequence boundaries are characterized by angular unconformities with a wide distribution. In well logging, the three conventional lithological curves above and below the sequence boundary of the volcanic eruption cycle exhibit abrupt jumps, which can be used as identification markers. In seismic logging, the erosion below the sequence boundary and the overburden above it are clearly defined, and the instantaneous phase plane properties extracted along the sequence boundary also show regular zonal variations.

[0058] Volcanic eruption sequences are controlled by the volcanic structure, primarily reflecting the spatial contact relationships of eruption products from different eruption phases separated by intermittent periods within the volcanic structure. The distribution of corresponding sequence boundaries between eruption sequences is typically confined to the area of ​​a single volcanic structure. In core samples, this manifests as transition surfaces of lithological assemblages; in well logging, it can be identified using combinations of bell-shaped and box-shaped, or bell-shaped and funnel-shaped logging curve envelopes, and these combinations exhibit spatial repeatability; in seismic data, significant differences in amplitude and frequency often exist above and below the interface.

[0059] Volcanic eruption rhythm reflects the spatial relationship between ejecta formed in the same eruption within the volcanic structure. The corresponding sequence boundary of volcanic eruption rhythm appears as a lithological assemblage transition surface in core samples. In well logging, it mainly appears as a gradient surface within a single envelope shape; in seismic logging, due to resolution limitations, it usually only shows differences in amplitude energy, making identification difficult.

[0060] Therefore, the inversion method for volcanic rocks in this embodiment mainly uses the volcanic eruption cycle and the sequence boundary of the volcanic eruption period for stepwise constraint.

[0061] like Figure 2 As shown, step S3, which uses the RBF algorithm to establish the initial model, includes the following steps performed sequentially:

[0062] a1) Apply the RBF neural network learning algorithm to preprocess the well data and seismic data respectively, and record the output results.

[0063] a2) Based on the output results obtained in step a1), the output error generated by each input vector is statistically analyzed. Combining the distribution of the error in the input space and the effect of each RBF on the network, adaptive RBF optimization prediction is performed.

[0064] a3) Based on the adaptive RBF optimization prediction results obtained in step a2), the results are compared with well-seismic data, and the joint loss function is calculated using the gradient descent training method. If the joint loss function does not converge, the process is restarted from step a2) until the results converge, and the initial model can be output.

[0065] In step a1), let the input vector x = (x1, x2, ... x n ) T With m hidden layer nodes, the output of the RBF network can be defined as: In the formula, ω i It is the weight between the i-th hidden layer node and the output layer. is the hidden layer activation function.

[0066] The following Gaussian function form is adopted: In the formula, σ i and ei These represent the width and center vector of the hidden layer node, respectively; ||xe i || represents the Euclidean norm.

[0067] According to existing technology, the more RBFs selected, the higher the approximation accuracy of the network, but the generalization ability of the network will decrease. Therefore, under the condition of satisfying a certain approximation accuracy, as few center vectors as possible should be selected to ensure that the network has a good generalization ability.

[0068] In this embodiment, based on the non-uniform distribution of the network's output error in the input space and the magnitude of the effect of each RBF on the network, the network parameters are adaptively adjusted through corresponding addition and deletion strategies, so that the network's approximation performance and generalization ability meet the requirements. At the same time, the network's training and operation can be carried out alternately, so it can adapt to the slow changes in the external environment.

[0069] Adding hidden nodes involves identifying points with large errors through comparison, and then appropriately inserting hidden nodes near these points. Let (x... k ,y k k = 1, 2, ..., N is a set of training samples. Initially, the number of hidden layer nodes is zero. Each time an addition operation is performed, the following criteria are used to determine whether to add a hidden layer node: In the formula It is the mean square error of the RBF network output, e k,nearest and x k,nearest Corresponding to the input vector x respectively k The closest hidden node center and input vector. If the addition condition is met, then (x k +x k,nearest Set ε / 2 as the new hidden node center, and set ε as the hidden node center. k Set the weight of the new node, and take the center width as...

[0070] Since the RBF neural network is a local perceptual field network, the overall output depends on the weights between the hidden and output layers and the distance between the hidden layer node center and the input vector. During training, the selected training samples are relatively sparse, while during the validation process after training, the validation data is generally denser. If some input vectors are close to the hidden layer center, the output will be significantly affected, thus worsening the network's generalization ability. Therefore, the removal of hidden layer nodes is introduced.

[0071] The deletion of hidden layer nodes is proposed based on the varying contributions of each hidden layer node to the entire network. Nodes with larger contributions are retained, while those with smaller contributions are deleted. For any hidden layer node s, P... sTo represent its contribution to the entire network, P s Defined as: Where, λ r y represents the proportion of the hidden unit r. r To standardize the weights, Ai is normalized before performing the deletion operation. like Then delete the s-th hidden node, where P max The maximum preset threshold for nodes is θ, where θ is the decision threshold.

[0072] The adaptive RBF optimization prediction results obtained above are used to calculate the joint loss function through well-seismic joint comparison. The loss function is defined as follows: Where, x n Input values ​​for the sample points, t n This represents the output of the corresponding sample in layer k, where N is the number of sample points, N0 is the number of hidden unit layers, and N... kH Let ω be the number of hidden units in the k-th layer. jk φ is the output weight of sample j in the k-th layer. j (x n Let W be an implicit unit function, where n, i, j, and k are indexed variables. The error surface function is obtained by calculating the component derivatives of W. The vector with the fastest ascent along the error surface is called the gradient of the error function W with respect to a certain component, denoted as [equation missing]. The relative width of W and its weight ω ji Negative gradient:

[0073]

[0074] Where c i σ is the center of the i-th Gaussian component. i Let L be the radius of the i-th Gaussian component, and L and U be the space where the sample is located.

[0075] The adjustment methods for the width and weights of the radial basis function along the gradient descent direction are as follows:

[0076]

[0077] In the formula, η and β∈[0,1] are the learning efficiencies of the width and weight, respectively, which determine the gradient search step size; t is the number of iterations.

[0078] When the calculated joint loss function converges, the initial model can be output.

[0079] like Figure 3 As shown, Figure 3 (a) is the target layer impedance plane diagram of the low-frequency model established by conventional well interpolation. Figure 3(b) is the impedance plane diagram of the target section under sequence constraint RBF modeling. The impedance model plane properties obtained by the two methods show significant differences in the high-low transition zone and areas with low well control. The impedance model obtained by RBF modeling in this invention is more consistent with the structural trend and has a better extrapolation effect. Its reservoir prediction results are consistent with the macro-geological understanding of the Junggar Basin and have a high degree of matching with the single-well logging integrated interpretation model, providing seismic geological basis for increasing the reserve and production of unconventional reservoirs in the Junggar Basin.

[0080] In this embodiment, to complete phase control inversion, it is first necessary to clarify the logging and seismic response characteristics of different volcanic rock facies, and conduct comparative analysis of various attributes such as amplitude, continuity, and frequency. Attributes sensitive to volcanic rock facies are selected, and multi-attribute fusion is carried out to reflect the characteristics of different volcanic rocks, generating a multi-attribute fusion data volume. Then, using the multi-attribute fusion data volume as input, phase control inversion is performed under the constraint of volcanic rock sequence to predict high-quality volcanic rock reservoirs.

[0081] like Figure 4 As shown, Figure 4 (a) is the inversion diagram of the interpolated wave impedance in the well. Figure 4 (b) is the RBF inversion diagram under sequence constraints. As can be seen from the diagram, the wave impedance inversion results based on well interpolation are heavily modeled and cannot reflect the lateral variation law of Carboniferous volcanic rock facies. The RBF inversion results based on sequence constraints can better reflect the macroscopic distribution law of volcanic structure, the variation relationship of different volcanic rock facies zones is clearer, and the characteristics of volcanic rock bodies are clear.

Claims

1. A RBF inversion method under special reservoir sequence constraints, characterized in that, The method includes the following steps performed sequentially: S1. Well logging curve preprocessing: Perform well logging curve environmental correction, multi-well consistency processing and multi-well standardization processing on the original well logging curves to obtain preprocessed well logging curves and well data, determine the target layer in the study area, and statistically analyze the VSP formation velocity. The VSP refers to the vertical seismic profile. Fault focusing processing: interpretive preprocessing of seismic data, filtering out high-frequency components that are not fault features at fault locations to obtain seismic data with improved accuracy; Fine calibration of wells and seismic data was performed using pre-processed well logging curves and improved seismic data. S2. Based on the preprocessed well logging curves and the improved seismic data, the volcanic eruption cycles and the sequence boundaries of the volcanic eruption periods are constrained step by step. At the same time, based on the fine calibration of wells and seismic data, the volcanic rock sedimentary discontinuities are identified, the volcanic rock sequence boundaries are divided, and the stratigraphic framework is constructed. S3. Based on the established stratigraphic framework, an initial model is built using the RBF algorithm based on well data and seismic data; S4. Using the multi-attribute fusion data volume as input, a low-frequency model is established under the constraint of volcanic rock sequence based on the obtained initial model, and phase control inversion is carried out to predict high-quality volcanic rock reservoirs.

2. The method of claim 1, wherein, Step S3 includes the following steps performed sequentially: a1) Apply the RBF neural network learning algorithm to preprocess well data and seismic data respectively, and record the output results; a2) Based on the output results obtained in step a1), calculate the output error generated by each input vector, and combine the distribution of the error in the input space with the effect of each RBF on the network to perform adaptive RBF optimization prediction; a3) Based on the adaptive RBF optimization prediction results obtained in step a2), the results are compared with well-seismic data, and the joint loss function is calculated using the gradient descent training method. If the joint loss function does not converge, the process is restarted from step a2) until the results converge, and the initial model can be output.

3. The method of claim 2, wherein, The joint loss function in step a3) is defined as follows: , In the formula, Enter values ​​for the sample points. For the corresponding sample in The output of the layer, The number of sample points The number of hidden unit layers. For the first The number of hidden unit layers in a layer. For the first layer The output weights of the sample points For implicit unit functions, , , , All are indexed variables; relative width and weight Negative gradient: ; in For the first The center of each Gaussian component For the first The radii of the Gaussian components, where L and U are the space where the sample is located; The adjustment methods for the width and weights of the radial basis function along the gradient descent direction are as follows: , In the formula, , are the learning efficiency of width and weight, respectively, determining the gradient search step size; is the number of iterations.