Automatic surrounding rock label optimization method based on pseudo label weighting and belief propagation

Abstract

The invention relates to the technical field of surrounding rock label automatic optimization, and discloses a surrounding rock label automatic optimization method based on pseudo label weighting and belief propagation. The objective of the invention is to solve the problem that spatial correlation and feature coupling among data are not fully utilized in the prior art; a differential processing mechanism for samples with different confidence degrees is lacked; an iterative optimization closed-loop system is not formed; and fuzziness between adjacent levels cannot be effectively processed. The method comprises the following steps: S1, data preparation and preprocessing; s2, initial model training; s3, confidence evaluation and sample classification; s4, adjusting labels and weights; s5, iterative optimization is carried out; and S6, outputting a final model. Through multi-round iterative optimization, the label quality is gradually improved, and finally the accuracy and reliability of the surrounding rock stability identification model are improved.

Description

Technical Field

[0001] This invention relates to the field of automatic optimization technology for surrounding rock labels, and in particular to an automatic optimization method for surrounding rock labels based on pseudo-label weighting and confidence propagation. Background Technology

[0002] With the rapid development of tunnel boring machine (TBM) technology, sensor systems can collect multi-source heterogeneous data in real time during the tunneling process, including key parameters such as total thrust, propulsion speed, and cutterhead torque. These parameters are closely related to the stability of the surrounding geological rock, providing a data foundation for intelligent geological identification. However, in practical engineering applications, rock stability labeling mainly relies on manual experience or indirect measurement methods, which suffers from problems such as high labeling costs, strong subjectivity, and poor consistency.

[0003] Traditional machine learning methods directly use these noisy labels for model training, which leads to decreased model performance and insufficient generalization ability. Although there are some label correction methods in the existing technology, such as confidence threshold filtering and multi-model ensemble voting, these methods have the following limitations: (1) they fail to make full use of the spatial correlation and feature coupling between data; (2) they lack a differentiated processing mechanism for samples with different confidence levels; (3) they do not form a closed-loop system for iterative optimization; and (4) they cannot effectively handle the ambiguity between adjacent levels.

[0004] Especially in the field of tunnel engineering, there is a strong continuity and transition between rock stability levels, and adjacent levels often exhibit similar characteristics. This makes it difficult for confidence assessment methods based solely on maximum probability to accurately identify the authenticity of labels. Therefore, an intelligent label optimization method is needed that can comprehensively consider confidence level, prediction consistency, and correlation between adjacent levels. Summary of the Invention

[0005] In view of the above technical problems, this disclosure provides an automatic optimization method for surrounding rock labels based on pseudo-label weighting and confidence propagation, which solves the technical problems in the prior art that fail to fully utilize the spatial correlation and feature coupling between data; lack a differentiated processing mechanism for samples with different confidence levels; do not form an iterative optimization closed-loop system; and cannot effectively handle the ambiguity between adjacent levels.

[0006] This invention presents a sample classification method based on multi-level confidence: samples are divided into four categories: most reliable, most suspicious, reliable, and suspicious, with different labeling and weighting strategies applied to each category. The method assesses and utilizes the proximity of confidence levels between adjacent levels: by calculating the difference between the highest and second-highest probabilities, ambiguity between adjacent levels is identified, and sample weights are adjusted accordingly. An iterative label optimization mechanism is implemented: label quality is progressively optimized through multiple iterations, with labels and weights adjusted in each round based on the current model's prediction confidence. A loss function is designed to address ambiguity between adjacent categories: an adjacent category constraint term is added to the standard cross-entropy loss, allowing the model to better reflect the continuous changes in surrounding rock stability. An overall confidence index is designed and applied: a composite index combining the highest probability value and the probability difference is used to evaluate overall label quality and determine the timing of iteration termination. Feature engineering methods for TBM tunneling data are included: methods specifically designed for the characteristics of tunneling data, such as temporal feature extraction, outlier handling, and missing value imputation.

[0007] According to one aspect of this disclosure, an automatic optimization method for surrounding rock labels based on pseudo-label weighting and confidence propagation is provided, comprising the following steps:

[0008] S1: Data preparation and preprocessing, cleaning and standardizing the TBM tunneling time series data, including outlier removal, missing value imputation, and feature standardization, to form a normalized dataset X and an initial noise label set Y0;

[0009] S2: Initial model training. The weighted LightGBM classifier is used for initial training on the dataset (X,Y0), with all initial weights set to 1.

[0010] S3: Confidence assessment and sample classification. The trained model is used to predict the probability distribution of each sample on the training set, resulting in P = [p1, p2, ..., p...]. i ], where pᵢ represents the probability that a sample belongs to the i-th class; calculated based on the probability distribution.

[0011] Maximum probability value: ;

[0012] Difference in probability between adjacent levels: ,in It is the second most probable value;

[0013] Based on the set high confidence threshold Threshold for probability difference between adjacent levels The samples are divided into four categories: most credible samples, most suspicious samples, credible samples, and suspicious samples.

[0014] S4: Label and weight adjustment, adopting different label and weight strategies for samples of different categories:

[0015] The most reliable sample retains its original label and is given a high weight. ;

[0016] The most suspicious sample will have its label changed to the predicted label and assigned a low weight. ;

[0017] Trustworthy samples retain their original labels and are assigned moderate weights. ;

[0018] Suspicious samples retain their original labels and are assigned low weights. ;

[0019] S5: Iterative optimization. Retrain the model using the adjusted labels and weights, repeating steps S3-S4 until the overall confidence index C no longer increases and the maximum number of iterations is reached. Termination condition where accuracy or F1 score on the validation set no longer improves;

[0020] S6: Final model output. Select the model that performs best on the validation set during the iteration process as the final output.

[0021] In some embodiments of this disclosure, in S3, the high confidence threshold The value range is 0.7-0.9, which is the threshold for the probability difference between adjacent levels. The value range is 0.1-0.3.

[0022] In some embodiments of this disclosure, the rules for sample classification in step S3 are as follows:

[0023] If the predicted label is consistent with the original label, then the sample is the most reliable sample;

[0024] However, if the predicted label is inconsistent with the original label, the sample is the most suspicious sample;

[0025] and If so, then the sample is a reliable sample;

[0026] and If so, the sample is a suspicious sample.

[0027] In some embodiments of this disclosure, the high weight in S4 is 1.0, the low weight is 0.1, and the medium weight is 0.7.

[0028] In some embodiments of this disclosure, the formula for calculating the overall confidence index C in step S5 is as follows: In the formula, N is the sample size and α is the balance coefficient. Let i be the maximum predicted probability value for the i-th sample. Let be the probability difference between adjacent grades of the i-th sample.

[0029] In some embodiments of this disclosure, the data preparation and preprocessing process in S1 further includes a feature extraction step: using a sliding window technique to process time-series data and extract statistical features, calculated using the following formula:

[0030] Moving average:

[0031] Moving variance:

[0032] Rate of change:

[0033] Where W is the window size. x t Let be the eigenvalue at time t.

[0034] In some embodiments of this disclosure, the initial model training of S2 further includes a weighted classification step: using LightGBM as the base classifier to adjust sample weights, with the objective function being a weighted multi-class cross-entropy.

[0035]

[0036] in The weights of sample i, Let i be the label (one-hot encoded) of class j to which sample i belongs. Predict the probability that sample i belongs to class j for the model.

[0037] In some embodiments of this disclosure, the confidence assessment in S3 includes the following steps: calculating a confidence index for each sample:

[0038] Main confidence level: ;

[0039] Ambiguity index: ;

[0040] Distribution uniformity: .

[0041] In some embodiments of this disclosure, the degree of difference between the model's predicted value and the true value in S3 is calculated using the following loss function:

[0042] The loss function is a weighted multi-class cross-entropy loss with a regularization term added:

[0043]

[0044] in:

[0045] The weight of sample i is determined based on the confidence category;

[0046] The true label for sample i after multiple rounds of optimization;

[0047] Predict the probability that sample i belongs to class j for the model;

[0048] This is a regularization term for model complexity.

[0049] The regularization coefficient is used.

[0050] Add a constraint term to the loss function to account for the ambiguity of adjacent categories:

[0051]

[0052] in:

[0053] d(j,k) is the neighbor distance function between categories j and k:

[0054]

[0055] β is the constraint strength coefficient for adjacent categories;

[0056] This loss condition encourages the model to produce similar predicted probabilities for adjacent categories, thus reflecting the continuity between surrounding rock stability levels.

[0057] The final complete loss function is: .

[0058] In some embodiments of this disclosure, the TBM tunneling time series data in S1 includes timestamps, ring numbers, mileage, total thrust, propulsion speed, cutterhead torque, cutterhead rotation speed, penetration depth, shield displacement, shield pressure, and initial surrounding rock stability labels.

[0059] In some embodiments of this disclosure, the preprocessing in S1 includes the following steps:

[0060] Outlier handling: Outliers are identified and corrected using the 3σ principle.

[0061]

[0062] Missing value handling: Spatiotemporal bilinear interpolation method is used.

[0063]

[0064] Where d t Let be the mileage value at time t;

[0065] Feature standardization: Z-score standardization is used.

[0066]

[0067] Where μ is the mean of the feature and σ is the standard deviation of the feature;

[0068] Feature engineering: generating derived features

[0069] Characteristics of the rate of change: ;

[0070] Acceleration characteristics: .

[0071] In some embodiments of this disclosure, the final model output parameters in step S6 include:

[0072] Optimized surrounding rock stability labels: ;

[0073] Final confidence level for each sample: ;

[0074] The trained surrounding rock stability recognition model: ;

[0075] Weight distribution of each sample: .

[0076] The beneficial effects of this invention are as follows:

[0077] 1. High degree of automation: Noise labels can be automatically corrected without human intervention, greatly reducing the cost of manual labeling.

[0078] 2. High adaptability: It can handle tag noise with different noise levels and distribution characteristics and has strong robustness.

[0079] 3. Clear physical meaning: It fully considers the continuity and transition between the stability levels of the surrounding rock, and conforms to the actual engineering situation.

[0080] 4. Good convergence: Through multiple rounds of iterative optimization, the label quality and model performance can be gradually improved.

[0081] 5. High interpretability: The confidence-based sample classification strategy has clear decision-making logic and interpretability.

[0082] 6. Excellent overall performance: Simultaneously optimizes model accuracy, confidence level and label quality, achieving multi-objective optimization. Attached Figure Description

[0083] Figure 1Flowchart of an automatic optimization method for surrounding rock labels based on pseudo-label weighting and confidence propagation; Detailed Implementation

[0084] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0085] Example 1

[0086] This example discloses an automatic optimization method for surrounding rock labels based on pseudo-label weighting and confidence propagation. See [link to relevant documentation]. Figure 1 The steps include:

[0087] S1: Data preparation and preprocessing, cleaning and standardizing the TBM tunneling time series data, including outlier removal, missing value imputation, and feature standardization, to form a normalized dataset X and an initial noise label set Y0;

[0088] S2: Initial model training. The weighted LightGBM classifier is used for initial training on the dataset (X,Y0), with all initial weights set to 1.

[0089] S3: Confidence assessment and sample classification. The trained model is used to predict the probability distribution of each sample on the training set, resulting in P = [p1, p2, ..., p...]. i ], where pᵢ represents the probability that a sample belongs to the i-th class; calculated based on the probability distribution.

[0090] Maximum probability value: ;

[0091] Difference in probability between adjacent levels: ,in It is the second most probable value;

[0092] Based on the set high confidence threshold Threshold for probability difference between adjacent levels The samples are divided into four categories: most credible samples, most suspicious samples, credible samples, and suspicious samples.

[0093] S4: Label and weight adjustment, adopting different label and weight strategies for samples of different categories:

[0094] The most reliable sample retains its original label and is given a high weight. ;

[0095] The most suspicious sample will have its label changed to the predicted label and assigned a low weight. ;

[0096] Trustworthy samples retain their original labels and are assigned moderate weights. ;

[0097] Suspicious samples retain their original labels and are assigned low weights. ;

[0098] S5: Iterative optimization. Retrain the model using the adjusted labels and weights, repeating steps S3-S4 until the overall confidence index C no longer increases and the maximum number of iterations is reached. Termination condition where accuracy or F1 score on the validation set no longer improves;

[0099] S6: Final model output. Select the model that performs best on the validation set during the iteration process as the final output.

[0100] In S3, the high confidence threshold The value range is 0.7-0.9, which is the threshold for the probability difference between adjacent levels. The value range is 0.1-0.3.

[0101] In S3, the rules for sample classification are as follows:

[0102] If the predicted label is consistent with the original label, then the sample is the most reliable sample;

[0103] However, if the predicted label is inconsistent with the original label, the sample is the most suspicious sample;

[0104] and If so, then the sample is a reliable sample;

[0105] and If so, the sample is a suspicious sample.

[0106] In S4, the high weight is set to 1.0, the low weight to 0.1, and the medium weight to 0.7.

[0107] The formula for calculating the overall confidence index C in S5 is as follows: In the formula, N is the sample size and α is the balance coefficient. Let i be the maximum predicted probability value for the i-th sample. Let be the probability difference between adjacent grades of the i-th sample.

[0108] The data preparation and preprocessing process in S1 also includes a feature extraction step: using the sliding window technique to process time-series data, extracting statistical features, and the calculation formula is as follows:

[0109] Moving average:

[0110] Moving variance:

[0111] Rate of change:

[0112] Where W is the window size. x t Let be the eigenvalue at time t.

[0113] The initial training of the S2 model also includes a weighted classification step: LightGBM is used as the base classifier to adjust the sample weights, and the objective function is a weighted multi-class cross-entropy.

[0114]

[0115] in The weights of sample i, Let i be the label (one-hot encoded) of class j to which sample i belongs. Predict the probability that sample i belongs to class j for the model.

[0116] The confidence assessment in S3 includes the following steps: calculating the confidence index for each sample:

[0117] Main confidence level: ;

[0118] Ambiguity index: ;

[0119] Distribution uniformity: .

[0120] The degree of difference between the model's predicted values ​​and the actual values ​​in S3 is calculated using the following loss function:

[0121] The loss function is a weighted multi-class cross-entropy loss with a regularization term added:

[0122]

[0123] in:

[0124] The weight of sample i is determined based on the confidence category;

[0125] The true label for sample i after multiple rounds of optimization;

[0126] Predict the probability that sample i belongs to class j for the model;

[0127] This is a regularization term for model complexity.

[0128] The regularization coefficient is used.

[0129] Add a constraint term to the loss function to account for the ambiguity of adjacent categories:

[0130]

[0131] in:

[0132] d(j,k) is the neighbor distance function between categories j and k:

[0133]

[0134] β is the constraint strength coefficient for adjacent categories;

[0135] This loss condition encourages the model to produce similar predicted probabilities for adjacent categories, thus reflecting the continuity between surrounding rock stability levels.

[0136] The final complete loss function is: .

[0137] The TBM tunneling time series data in S1 includes timestamp, ring number, mileage, total thrust, propulsion speed, cutterhead torque, cutterhead rotation speed, penetration depth, shield displacement, shield pressure, and initial surrounding rock stability label.

[0138] The preprocessing in S1 includes the following steps:

[0139] Outlier handling: Outliers are identified and corrected using the 3σ principle.

[0140]

[0141] Missing value handling: Spatiotemporal bilinear interpolation method is used.

[0142]

[0143] Where d t Let be the mileage value at time t;

[0144] Feature standardization: Z-score standardization is used.

[0145]

[0146] Where μ is the mean of the feature and σ is the standard deviation of the feature;

[0147] Feature engineering: generating derived features

[0148] Characteristics of the rate of change: ;

[0149] Acceleration characteristics: .

[0150] The final model output parameters in S6 include:

[0151] Optimized surrounding rock stability labels: ;

[0152] Final confidence level for each sample: ;

[0153] The trained surrounding rock stability recognition model: ;

[0154] Weight distribution of each sample: .

[0155] Although some preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0156] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for automatic optimization of surrounding rock labels based on pseudo-label weighting and confidence propagation, characterized in that, Comprising the following steps: S1: data preparation and preprocessing, cleaning and standardizing the TBM tunneling time series data, including outlier removal, missing value interpolation, feature standardization, forming a standardized data set X and an initial noise label set Y0; S2: initial model training, using a weighted LightGBM classifier to perform initial training on the data set (X, Y0), and setting all initial weights to 1; S3: Confidence evaluation and sample classification, using the trained model to predict the training set to get the prediction probability distribution P = [p1, p2,..., p i ] of each sample, where pᵢ represents the probability that the sample belongs to the i-th class; Based on the probability distribution calculation: the maximum probability value: ; adjacent level probability difference: , wherein is the second largest probability value; according to the set high confidence threshold and the adjacent level probability difference threshold , the sample is divided into four categories: the most credible sample, the most suspicious sample, the credible sample and the suspicious sample. S4: Label and weight adjustment, different label and weight strategies are taken for different categories of samples: the most credible samples keep the original label and are given high weight ; the most suspicious samples change the label to the predicted label and are given low weight ; the credible samples keep the original label and are given medium weight ; the suspicious samples keep the original label and are given low weight ; S5: Iterative optimization, retrain the model using the adjusted labels and weights, repeat S3-S4 steps until the overall confidence metric C no longer improves, reaches a maximum number of iterations a termination condition when the accuracy or F1 score on the validation set no longer improves S6: final model output, selecting the model with the best performance on the validation set in the iteration process as the final output.

2. The pseudo-label weighting and belief propagation based automatic optimization method for surrounding rock labels according to claim 1, characterized in that: The high confidence threshold in the S3 The value range of the adjacent level probability difference threshold The value range of the adjacent level probability difference threshold The rule for sample classification in S3 is: and the predicted label is consistent with the original label, the sample is the most credible sample; but the predicted label is inconsistent with the original label, the sample is the most suspicious sample; and the sample is a credible sample; and the sample is a suspicious sample.

3. The method of claim 1, wherein the method is characterized by: The high weight in S4 is 1.0, the low weight is 0.1, and the medium weight is 0.7; The calculation formula of the overall confidence index C in S5 is: , wherein N is the sample quantity, a is a balance coefficient, is the maximum prediction probability value of the i th sample, is the adjacent grade probability difference of the i th sample.

4. The method of claim 1, wherein the method is based on pseudo label weighting and confidence propagation for automatic optimization of surrounding rock labels. The data preparation and preprocessing process in the S1 further includes the step of feature extraction: using sliding window technology to process time series data, extracting statistical features, and the calculation formula is as follows: moving average: ; Moving variance: Rate of change: Where W is the window size, x t is the feature value at time t.

5. The method of claim 1, wherein the method is characterized by: The S2 initial model training further comprises the step of weighted classification: using LightGBM as the basic classifier for adjusting the sample weight, and the objective function is weighted multi-class cross-entropy: where is the weight of sample i, is the label (one-hot encoding) of sample i belonging to class j, is the probability that the model predicts sample i belongs to class j.

6. The pseudo-label weighting and belief propagation based automatic optimization method for surrounding rock labels according to claim 1, characterized in that: The confidence evaluation in S3 includes the following steps: calculating the confidence indicator of each sample: main confidence: ; fuzziness indicator: ; distribution uniformity: .

7. The method of claim 1, wherein the method is based on pseudo label weighting and confidence propagation for automatic optimization of surrounding rock labels. The difference between the predicted value and the true value of the model in S3 is calculated by the following loss function: The loss function is a weighted multi-class cross-entropy loss with a regularization term: Wherein: the weight for sample i is determined according to the confidence class; yi is the true label for sample i after multiple rounds of optimization; pi,j is the probability that the model predicts that sample i belongs to class j; is the model complexity regularization term; is the regularization coefficient; A constraint term for adjacent category ambiguity is added to the loss function: Wherein: d(j, k) is the adjacent distance function between classes j and k, ; β is the adjacent class constraint strength coefficient; The final complete loss function is: .

8. The method of claim 1, wherein the method is based on pseudo label weighting and confidence propagation for automatic optimization of surrounding rock labels. The TBM tunneling time series data in S1 includes timestamp, ring number, mileage, total thrust, propulsion speed, cutter torque, cutter speed, penetration, shield displacement, shield pressure, and initial surrounding rock stability label; The preprocessing in S1 comprises the following steps: a. Outlier processing: 3σ principle is adopted to identify and correct outliers: b. Missing value processing: spatio-temporal bilinear interpolation method is adopted: where d t is the distance value at time t; c. Feature standardization: Z-score standardization is adopted: Wherein μ is the mean of the feature, and σ is the standard deviation of the feature; d. Feature engineering: Generate derived features; rate of change features: ; acceleration features: .

9. The method of claim 1, wherein the method is based on pseudo label weighting and confidence propagation for automatic optimization of surrounding rock labels. The final model output parameter in the S6 includes: the optimized surrounding rock stability label: ; the final confidence of each sample: ; the trained surrounding rock stability identification model: ; the weight distribution of each sample: .

Images