Drilling infrared spectrum missing data completion method

Abstract

The application relates to a drilling infrared spectrum missing data completion method, and belongs to the field of data processing.The method comprises the following steps: obtaining a training data set, wherein samples in the training data set are drilling infrared spectrum missing data, the drilling infrared spectrum missing data has complete short-wave infrared spectrum data and missing thermal infrared spectrum data; a drilling infrared spectrum missing data completion model is trained based on the training data set; drilling infrared spectrum missing data to be processed is input into the trained model to obtain a prediction result of thermal infrared spectrum data.On the basis of no classifier guidance, firstly, the conditional guidance residual is subjected to frequency domain decomposition, then, according to recovery targets in different stages of diffusion reverse sampling, low-frequency structural information and high-frequency detailed information are subjected to stage-based differential guidance, so that the model can give consideration to stable recovery of a thermal infrared spectrum overall energy baseline and fine reconstruction of local diagnostic absorption characteristics, and the authenticity, continuity and geological interpretability of the reconstructed spectrum are improved.

Description

Technical Field

[0001] This application relates to the field of data processing, and more specifically, to a method for completing missing borehole infrared spectral data. Background Technology

[0002] In mineral resource evaluation, fine mineral logging, and exploration borehole logging, short-wave infrared (SWIR) (wavelength <2500nm) and thermal infrared (TIR) ​​(wavelength ≥2500nm) hyperspectral techniques are core technologies for identifying alteration mineral communities, inverting mineral composition, and interpreting rock strata. Minerals exhibit diagnostic characteristic absorption peaks in specific infrared bands; for example, silicate minerals show residual radiation absorption peaks in the thermal infrared region. By analyzing the morphology and absorption positions of reflectance curves, high-precision mineral identification can be achieved.

[0003] However, in actual field logging and intensive indoor measurements, many uncertainties can affect the spectral data, resulting in serious incompleteness. These uncertainties include: complex underground physical environment, dust interference, sensor overheating caused by long-term high-load continuous operation of instruments, atmospheric window absorption, and loss of data transmission due to heterogeneous data.

[0004] The most typical manifestations of this incompleteness are the complete absence of an entire SWIR or TIR band, breaks in the spectral curve at key junctions, and local reflectance saturation anomalies. Thermal infrared spectroscopy, due to its high detector cooling requirements and fragile technical environment, is particularly prone to complete band loss. This directly leads to a significant reduction in the accuracy of subsequent automatic lithological logging, alteration mapping, and quantitative mineral inversion.

[0005] For the completion and reconstruction of missing segments in the geological spectrum, the existing conventional technical approaches have the following significant drawbacks: 1. Traditional interpolation and low-rank completion methods are difficult to handle the case of complete missing segments of the entire spectrum; 2. One-dimensional deep repair networks and generative adversarial networks (GANs) lack effective neighborhood context; 3. Remote sensing hyperspectral models and traditional generative models separate the multi-source prior information of geology; 4. The unified guidance of traditional denoising diffusion models across the entire frequency band leads to mutual interference between the "overall envelope and local details". Summary of the Invention

[0006] To overcome at least one deficiency in the prior art, this application provides a method for completing missing borehole infrared spectral data.

[0007] In one embodiment, a method for completing missing borehole infrared spectral data includes: Step S1: Obtain the training dataset. The samples in the training dataset are missing borehole infrared spectral data. The missing borehole infrared spectral data has complete shortwave infrared spectral data and missing thermal infrared spectral data. Step S2: Construct a model to complete missing borehole infrared spectral data. The model includes a feature embedding module and a diffusion network. Step S3: Train the model to complete missing borehole infrared spectral data based on the training dataset to obtain the trained model; During training, the input data passes through the feature embedding module to generate conditional control vectors; the diffusion network performs noise addition and denoising processes. The denoising process includes a first stage and a second stage. At each time step of the denoising process, conditional noise and unconditional noise are generated based on the conditional control vector using a classifier-free guidance mechanism. The classifier-free guidance conditional residual between the conditional noise and the unconditional noise is calculated. A real-valued Fast Fourier Transform (FFT) is performed on the classifier-free guidance conditional residual to obtain the Fourier transform result. The Fourier transform result is multiplied by a low-frequency mask and then subjected to an inverse Fast Fourier Transform (IFFT) to obtain the low-frequency conditional residual projection. The Fourier transform result is multiplied by a high-frequency mask and then subjected to an IFFT to obtain the high-frequency conditional residual projection. For each time step of the first stage, the predicted noise for each time step of the first stage is generated based on the conditional noise and the low-frequency conditional residual projection. For each time step of the second stage, the predicted noise for each time step of the second stage is generated based on the conditional noise and the high-frequency conditional residual projection. Step S4: Input the missing borehole infrared spectrum data to be processed into the trained model to obtain the prediction results of the thermal infrared spectrum data.

[0008] In one embodiment, the predicted noise for each time step of the first stage is generated based on the conditional noise and the low-frequency conditional residual projection, using the following formula:

[0009] in, For the first phase of time steps Predictive noise, For the first phase of time steps Conditional noise, For time steps Low-frequency conditional residual projection, This is the low-frequency guidance enhancement factor.

[0010] In one embodiment, the predicted noise for each time step of the second stage is generated based on the conditional noise and the projection of the high-frequency conditional residual, using the following formula:

[0011] in, For the second phase of time steps Predictive noise, For the second phase of time steps Conditional noise, For time steps High-frequency conditional residual projection, This represents the high-frequency guidance enhancement factor.

[0012] In one embodiment, generating conditional noise and unconditional noise based on a classifier-free guidance mechanism using a conditional control vector includes: The conditional control vector is input into the diffusion network; For each sample in each training batch, a probability value is randomly generated. If the probability value is less than the preset discard probability, the conditional control vector corresponding to the sample and input into the diffusion network is replaced with an empty conditional vector with all zeros. Otherwise, no replacement is made. The diffusion network generates conditional noise based on the conditional control vector and unconditional noise based on the all-zero empty conditional vector.

[0013] In one embodiment, the feature embedding module includes a one-dimensional residual convolutional neural network, an embedding layer, a position encoding layer, and a fusion module; the one-dimensional residual convolutional neural network, the embedding layer, and the position encoding layer are configured in parallel. The input data is processed through a one-dimensional residual convolutional neural network, an embedding layer, and a positional encoding layer to obtain semantically dense feature vectors, lithological feature vectors, and depth feature vectors, respectively. Semantic dense feature vectors, lithological feature vectors, and depth feature vectors are all input into the fusion module for concatenation to obtain the conditional control vector.

[0014] Compared with existing technologies, this application has the following advantages: The borehole infrared spectral missing data completion method of this application, without classifier guidance, first performs frequency domain decomposition on the condition-guided residual, and then implements phased differentiated guidance on low-frequency structural information and high-frequency detail information according to the recovery targets of different stages of diffusion backsampling. This enables the model to take into account both the stable recovery of the overall energy baseline of the thermal infrared spectrum and the fine reconstruction of local diagnostic absorption features, thereby improving the authenticity, continuity and geological interpretability of the reconstructed spectrum. This application makes full use of the cross-modal correlation between shortwave infrared spectrum and thermal infrared spectrum, and achieves continuous reconstruction of the entire missing thermal infrared spectrum when only borehole shortwave infrared spectrum and related geological condition information are input, thereby improving the ability to reconstruct large-span cross-band thermal infrared spectra. Attached Figure Description

[0015] This application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which, together with the detailed description below, are incorporated in and form part of this specification. In the drawings: Figure 1 A flowchart of a method for completing missing borehole infrared spectral data is shown. Figure 2 A schematic diagram of a borehole infrared spectral data completion model is shown. Figure 3 A scatter plot of predicted and completed values ​​is shown; Figure 4 The histogram of the prediction error distribution is shown. Figure 5 The root mean square error histograms for different lithological types are shown. Figure 6 A visualization comparison of the thermal infrared hyperspectral reconstruction effects on test samples of different lithologies is shown. Detailed Implementation

[0016] Exemplary embodiments of the present application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of the actual embodiments are described in the specification. However, it should be understood that many embodiment-specific decisions can be made in the development of any such actual embodiment to achieve the developer’s specific objectives, and these decisions may vary as the embodiments differ.

[0017] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the device structure closely related to the solution of this application is shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0018] It should be understood that this application is not limited to the described embodiments by virtue of the following description with reference to the accompanying drawings. In this document, embodiments may be combined with each other, features may be substituted or borrowed between different embodiments, and one or more features may be omitted in one embodiment, where feasible.

[0019] This application provides a method for completing missing borehole infrared spectral data. Figure 1 A flowchart illustrating the method for completing missing borehole infrared spectral data is shown. (See [link]) Figure 1 The method mainly includes the following steps: Step S1: Obtain the training dataset. The samples in the training dataset are missing borehole infrared spectral data. The missing borehole infrared spectral data has complete shortwave infrared spectral data and missing thermal infrared spectral data.

[0020] Here, for each sampling point of each borehole, the acquired infrared spectral data includes short-wave infrared spectral data and thermal infrared spectral data. Due to the high cooling requirements of its detector and the fragile technical environment, the probability of complete loss of thermal infrared data is particularly prominent.

[0021] The input data also includes the lithological category label and absolute logging depth for each sampling point.

[0022] Step S2: Construct a model to complete missing borehole infrared spectral data. The model includes a feature embedding module and a diffusion network. Figure 2 A schematic diagram of a model for completing missing borehole infrared spectral data is shown.

[0023] Specifically, the feature embedding module includes a one-dimensional residual convolutional neural network (ResNet1D), an embedding layer, a positional encoding layer, and a fusion module. The one-dimensional residual convolutional neural network (ResNet1D), the embedding layer, and the positional encoding layer are configured in parallel.

[0024] The input data is processed through a one-dimensional residual convolutional neural network, an embedding layer, and a positional encoding layer to obtain semantically dense feature vectors, lithological feature vectors, and depth feature vectors, respectively. These three feature vectors are then concatenated in a fusion module to obtain a conditional control vector.

[0025] Here, the shortwave infrared spectral data in the input data is processed by a one-dimensional residual convolutional neural network to perform adaptive multi-scale downsampling feature extraction, and the output is a semantically dense feature vector containing the high-frequency absorption features of shortwave infrared mineral functional groups.

[0026] The lithological category labels in the input data, such as garnet-bearing gneiss, magnetite quartzite, silicified potassic altered rocks, etc., are projected and transformed into a lithological feature vector reflecting the lithological source and symbiotic assemblage characteristics after being embedded in the layer.

[0027] The absolute cataloging depth of the sampling points in the input data is projected into a depth feature vector by a high-dimensional spatial absolute position projection through a position encoding layer.

[0028] Step S3: Train the model to complete the missing data of borehole infrared spectroscopy based on the training dataset to obtain the trained model.

[0029] During training, the input data passes through the feature embedding module to generate conditional control vectors; the diffusion network performs the noise addition and denoising processes, and the diffusion network adopts the traditional U-Net structure.

[0030] During the noise addition process, the total number of discrete denoising time steps is defined. During the training phase, the input samples are dimensionality reduced to obtain the spectral residual vector. The spectral residual vector is used as the initial state, and Gaussian white noise is added step by step at random sampling time steps to generate noisy latent variables for the denoising process.

[0031] Traditional diffusion models employ a uniform condition-guided strategy for simultaneous denoising across the entire frequency band, leading to mutual interference between the overall envelope and local details. This embodiment, leveraging the physical response characteristics of thermal infrared spectroscopy—the coexistence of an overall continuous spectral shape and local diagnostic absorption features—first decomposes the condition-guided residuals in the frequency domain without classifier guidance. Then, based on the recovery targets at different stages of diffusion backsampling, it implements phased differentiated guidance for low-frequency structural information and high-frequency detail information. This allows the model to simultaneously achieve stable recovery of the overall energy baseline of the thermal infrared spectrum and refined reconstruction of local diagnostic absorption features, improving the realism, continuity, and geological interpretability of the reconstructed spectrum.

[0032] Specifically, the denoising process includes a first stage and a second stage. In the first stage, the time step... Low-frequency structures are prioritized for recovery, and in the second stage time step High-frequency diagnostic features are meticulously recovered. The specific process is as follows: At each time step, firstly, conditional noise and unconditional noise are generated based on the conditional control vector using a classifier-free guidance mechanism, including: The conditional control vector is input into the diffusion network. For each sample in each training batch, a probability value is randomly generated. If the probability value is less than the preset discard probability (e.g., 0.15), the conditional control vector input into the diffusion network corresponding to the sample is replaced with an empty conditional vector (i.e., the elements in the vector are all 0). Otherwise, no replacement is made. The diffusion network generates conditional noise based on the conditional control vector. Generate unconditional noise based on the all-zero empty conditional vector. .

[0033] Then, calculate the conditional noise. and unconditional noise Classifier-free guided condition residuals , ; Then, the residuals of the classifier-free guided conditions. Perform a real-number Fast Fourier Transform (RFFT) to obtain the Fourier transform result. ; Then, the Fourier transform result With low-frequency mask Multiply, then perform inverse fast Fourier transform The low-frequency conditional residual projection is obtained. Fourier transform results and high-frequency mask Multiply, then perform inverse fast Fourier transform The high-frequency conditional residual projection is obtained. The following formula is used to express this:

[0034]

[0035] The construction process of low-frequency masks and high-frequency masks is as follows: ,in, for The Middle There are n elements, and the frequency domain length is given. Define the low-frequency proportional control parameter as ρ, where:

[0036] Constructing a low-frequency mask:

[0037] in, Indicates low-frequency mask The Middle The mask value of each element; And obtain the high-frequency mask through complementary relationships.

[0038]

[0039]

[0040] in, Represents high-frequency mask The Middle The mask value of each element; ensure:

[0041] The two spaces are orthogonal:

[0042] The two constitute a complementary frequency projection space.

[0043] Then, for the time step of the first stage, the complete conditional noise prediction is retained as the basic denoising term, and only the low-frequency conditional residual is denoised. Enhanced guidance is implemented, and the predicted noise for each time step in the first stage is generated based on the conditional noise and the projection of the low-frequency conditional residual, using the following formula:

[0044] in, For the first phase of time steps Predictive noise, For the first phase of time steps Conditional noise, For time steps Low-frequency conditional residual projection, For low-frequency guidance enhancement coefficient, the preferred value is... .

[0045] The above operations guide the denoising network to prioritize the recovery of the overall spectral envelope, the continuous radiation energy baseline, and the overall trend of reflectance variation, so that random noise quickly converges to a low-frequency structural space that conforms to the overall distribution law of the geological spectrum, establishing a stable macroscopic spectral framework and providing a reliable foundation for subsequent local detail recovery.

[0046] For each time step of the second stage, the predicted noise for each time step is generated based on the conditional noise and the projection of the high-frequency conditional residual. Only the high-frequency conditional residual is considered. Enhanced guidance shifts the focus of noise reduction from overall structure restoration to local detail reconstruction, using the following formula:

[0047] in, For the second phase of time steps Predictive noise, For the second phase of time steps Conditional noise, For time steps High-frequency conditional residual projection, λ is the high-frequency guidance enhancement factor. high =0.25~0.50, preferably 0.25.

[0048] The above operations focus on recovering high-frequency information such as local absorption valleys, narrow-band diagnostic absorption features, and subtle spectral line changes in the thermal infrared spectrum, and gradually complete the iterative update of latent variables by combining the diffusion model's backsampling state update formula. While maintaining the stability of the global structure, this fully releases the expressive power of local high-frequency information and effectively avoids the problems of over-smoothing of absorption features and loss of detail caused by traditional unified denoising strategies.

[0049] At each time step in the denoising process, the data processed at the current time step is denoised according to the predicted noise to obtain the prediction result of the current time step; the prediction result of the last time step in the second stage is the prediction result of the final thermal infrared spectral data.

[0050] Step S4 involves inputting the missing borehole infrared spectral data to the trained model to obtain the prediction results of the thermal infrared data, ultimately completing the missing borehole infrared spectral data. Here, the data input into the trained model also includes lithological category information and absolute logging depth for each sampling point.

[0051] To further verify the effectiveness of the method in this application, the following experimental analysis was conducted.

[0052] The generalization performance of the method in this application was statistically analyzed on the 100ZK9102 drilling dataset, which verified the prediction accuracy and stability of this application in actual drilling scenarios.

[0053] Figure 3 The plot shows a scatter plot of predicted and completed values. The horizontal axis represents measured reflectance, and the vertical axis represents completed reflectance. The red dashed line is the y=x reference line. The coefficient of determination is marked on the plot. The figure shows that the model predictions and measured values ​​are significantly positively correlated overall, and the model can fit the trend of spectral intensity changes well.

[0054] Figure 4 The histogram showing the distribution of prediction errors is presented. The horizontal axis represents the prediction error (the difference between the predicted and measured values), and the vertical axis represents the sample size (i.e., frequency). The errors in the figure are generally normally distributed and concentrated around 0, indicating that the model has no systematic bias and the error fluctuations are within a reasonable range.

[0055] Figure 5 The root mean square error (RMSE) histograms for different lithological types are shown, with the horizontal axis representing lithological type and the vertical axis representing RMSE value. The results indicate that the model has stable prediction accuracy for all lithological types, and the RMSE differences between different lithologies are mainly determined by the complexity of spectral features. No obvious lithological prediction failure problem was found.

[0056] Figure 6 Visual comparisons of thermal infrared hyperspectral reconstruction results on test samples of different lithologies are shown. In each sub-figure, the horizontal axis represents the TIR spectral band index after dimensionality reduction, and the vertical axis represents reflectance. The solid black line represents the actual spectral data obtained from borehole measurements, and the dashed red line represents the reconstructed spectrum obtained in this application. The filled area between the two curves represents the deviation range between the predicted spectrum and the actual spectrum. The corresponding root mean square error (RMSE) and coefficient of determination (CDO) are labeled below each sub-figure. The proposed method utilizes quantitative indicators such as spectral angle matching (SAM). Results show that the proposed method can effectively reproduce the overall spectral morphology and key characteristic bands of different lithologies. The reconstructed spectra are consistent with the overall trends of the true spectra, with deviations controlled within a reasonable range. It also demonstrates stable reconstruction capabilities for spectral data of different lithologies, validating the generalizability and effectiveness of this application.

[0057] Compared with the prior art, this application has the following technical effects: 1. By fully utilizing the cross-modal correlation between shortwave infrared spectrum and thermal infrared spectrum, continuous reconstruction of missing thermal infrared spectrum can be achieved by only inputting borehole shortwave infrared spectrum and related geological conditions information. This breaks through the limitation of traditional interpolation methods that are only applicable to local band completion and improves the ability to reconstruct thermal infrared spectrum over a large span of bands.

[0058] 2. To address the issue that the overall spectral shape recovery and the recovery of local diagnostic absorption features in thermal infrared spectroscopy have different recovery rules, a conditional residual hierarchical guidance mechanism is proposed. Based on the recovery targets at different stages of diffusion backsampling, differentiated conditional guidance is applied to low-frequency overall structural information and high-frequency local detail information, respectively. This achieves synergistic optimization of overall spectral shape recovery and local diagnostic absorption feature recovery, effectively reducing problems such as overall spectral shape drift, excessive smoothing of local details, and loss of diagnostic information, thereby improving the quality of thermal infrared spectral reconstruction.

[0059] 3. Rejecting blind, end-to-end data brute-force fitting, it innovatively couples deep spatial coding, refined lithological semantics, and hyperspectral features into a deep, latent spatial model. By using multi-source conditions as strong physical constraints in the diffusion process, the depth of the completed spectral lines matches the petrogenic physical mechanism of the current sampling layer, and the generated spectral lines match the actual mineral co-occurrence patterns.

[0060] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for completing missing borehole infrared spectral data, characterized in that, include: Step S1: Obtain the training dataset. The samples in the training dataset are missing borehole infrared spectral data. The missing borehole infrared spectral data has complete shortwave infrared spectral data and missing thermal infrared spectral data. Step S2: Construct a model to complete missing borehole infrared spectral data, the model including a feature embedding module and a diffusion network; Step S3: Train the borehole infrared spectrum missing data completion model based on the training dataset to obtain the trained model; During training, the input data passes through the feature embedding module to generate a conditional control vector; the diffusion network performs noise addition and denoising processes. The denoising process includes a first stage and a second stage. At each time step of the denoising process, conditional noise and unconditional noise are generated based on the conditional control vector using a classifier-free guidance mechanism. The classifier-free guidance conditional residual between the conditional noise and unconditional noise is calculated. A real-valued Fast Fourier Transform (FFT) is performed on the classifier-free guidance conditional residual to obtain the Fourier transform result. The Fourier transform result is multiplied by a low-frequency mask and then subjected to an inverse Fast Fourier Transform (IFFT) to obtain the low-frequency conditional residual projection. The Fourier transform result is multiplied by a high-frequency mask and then subjected to an IFFT to obtain the high-frequency conditional residual projection. For each time step of the first stage, the predicted noise for each time step of the first stage is generated based on the conditional noise and the low-frequency conditional residual projection. For each time step of the second stage, the predicted noise for each time step of the second stage is generated based on the conditional noise and the high-frequency conditional residual projection. Step S4: Input the missing borehole infrared spectrum data to be processed into the trained model to obtain the prediction results of the thermal infrared spectrum data.

2. The method as described in claim 1, characterized in that, The prediction noise for each time step in the first stage is generated based on the conditional noise and the projection of the low-frequency conditional residual, using the following formula: in, For the first phase of time steps Predictive noise, For the first phase of time steps Conditional noise, For time steps Low-frequency conditional residual projection, This is the low-frequency guidance enhancement factor.

3. The method as described in claim 1, characterized in that, The prediction noise for each time step in the second stage is generated based on the conditional noise and the projection of the high-frequency conditional residual, using the following formula: in, For the second phase of time steps Predictive noise, For the second phase of time steps Conditional noise, For time steps High-frequency conditional residual projection, This represents the high-frequency guidance enhancement factor.

4. The method as described in claim 1, characterized in that, The generation of conditional noise and unconditional noise based on the conditional control vector using a classifier-free guidance mechanism includes: The conditional control vector is input into the diffusion network; For each sample in each training batch, a probability value is randomly generated. If the probability value is less than the preset discard probability, the conditional control vector corresponding to the sample and input into the diffusion network is replaced with an empty conditional vector with all zeros. Otherwise, no replacement is made. The diffusion network generates conditional noise based on the conditional control vector and unconditional noise based on the all-zero empty conditional vector.

5. The method as described in claim 1, characterized in that, The feature embedding module includes a one-dimensional residual convolutional neural network, an embedding layer, a position encoding layer, and a fusion module; the one-dimensional residual convolutional neural network, the embedding layer, and the position encoding layer are configured in parallel. The input data is processed through a one-dimensional residual convolutional neural network, an embedding layer, and a positional encoding layer to obtain semantically dense feature vectors, lithological feature vectors, and depth feature vectors, respectively. The semantically dense feature vector, the lithological feature vector, and the depth feature vector are all input into the fusion module for splicing to obtain the conditional control vector.

Images