Geological diggability label self-generation method based on adaptive loss collaborative optimization and equipment thereof
The self-generation method of geological excavability tags through adaptive loss collaborative optimization solves the problems of low tag accuracy and high cost in TBM construction, and generates high-precision and reliable geological excavability tags that are suitable for intelligent TBM tunneling.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-10
- Publication Date
- 2026-04-10
AI Technical Summary
In existing technologies, TBM construction relies on rough geological survey reports for labeling, resulting in low label accuracy and high signal-to-noise ratio. Initialization prediction using other engineering or numerical models leads to significant label bias. Manual labeling is costly and highly subjective, making it difficult to obtain a large number of high-quality labels. Furthermore, unsupervised methods lack clear physical meaning and are difficult to correlate with engineering stability levels.
A self-generation method for geological excavability labels based on adaptive loss collaborative optimization is adopted. Through iterative optimization of adaptive loss function and base learning model, combined with high confidence screening and physical rule constraints, high-precision and high-reliability geological excavability labels are generated.
It enables the automatic generation of accurate and reliable geological excavability labels from unlabeled TBM tunneling data, improving the model's generalization ability and interpretability. The output results conform to engineering mechanisms and are suitable for intelligent TBM tunneling parameter setting.
Smart Images
Figure CN121834167A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of geological excavability label self-generation, and particularly relates to a geological excavability label self-generation method based on adaptive loss collaborative optimization and a device thereof. BACKGROUND
[0002] In the construction process of a full-face tunnel boring machine (TBM), accurately grasping the excavability of the geological rock mass in front is the key to ensuring safe and efficient construction. Traditionally, the stability of surrounding rock is evaluated by relying on pre-geological survey (geological survey) data, but the geological survey data has the following inherent defects: (1) the sampling points are sparse, and it is difficult to fully reflect the fine changes of complex geological bodies such as tunnels along the line, especially faults and fracture zones; (2) the cost is high, and it cannot be updated in real time during the excavation process. In contrast, the TBM is equipped with a large number of sensors that can collect full-dimensional time series data reflecting the interaction between the cutter head and the rock mass at a high frequency (such as every 2 seconds), including total thrust F (kN), advancing speed v (mm / min), cutter head torque T (kN·m), cutter head speed n (rpm), etc. These data sequences imply rich geological information and are a valuable data source for rock mass excavability evaluation. Using these data to intelligently evaluate and predict the geological excavability has become a research hotspot in the industry.
[0003] At present, using machine learning methods to mine geological information from excavation data faces the dilemma of "data enrichment and label scarcity". That is, there is a large amount of excavation data X, but there is a lack of accurate geological labels y corresponding to it. Existing solutions mainly include: (1) relying on rough geological survey reports for rough labeling, which has low label accuracy and high signal-to-noise ratio; (2) using other engineering or numerical models for initial prediction, which has significant bias in the label, or using manual labeling (which is costly and subjective), making it difficult to obtain a large number of high-quality labels for model training; (3) completely unsupervised methods (such as clustering), which lack clear physical meaning and are difficult to correspond to the stability level in engineering.
[0004] Therefore, developing a method that can use low-precision initial labels to iteratively generate high-precision and high-reliability geological excavability labels through algorithms is of great significance for realizing the intelligentization and adaptive control of TBM construction. SUMMARY
[0005] In view of the above technical problems, this disclosure provides a method and device for self-generating geological excavability labels based on adaptive loss co-optimization. It solves the technical problems of existing technologies that rely on rough geological survey reports for coarse labeling, resulting in low label accuracy and high signal-to-noise ratio; use other engineering or numerical models for initial prediction, resulting in significant label bias; or use manual labeling, which is costly and subjective, making it difficult to obtain a large number of high-quality labels for model training; and completely unsupervised methods, whose results lack clear physical meaning and are difficult to correspond to engineering stability levels.
[0006] According to one aspect of this disclosure, a method for self-generating geological excavability labels based on adaptive loss collaborative optimization is provided, comprising the following steps: Step S1: Data preparation and initialization: Input an unlabeled TBM tunneling time series dataset. The D contains N samples, denoted as In the formula Given a d-dimensional feature vector, initial geological labels are generated for the dataset D through random assignment or by utilizing a weak engineering model. In the formula Provide discrete level labels or continuous value labels; initialize the adaptive loss function. Base learning model ; Step S2: Model Training: In the first step... In the next iteration, the current dataset is used. Current loss function The base learning model in training step S1 The training objective is to minimize the loss function. In the formula For the model parameter; Step S3: Prediction and Confidence Generation: Using the trained model For dataset Make predictions and obtain the prediction output for each sample. The probability distribution for each category, through Calculate the confidence score for each sample. and prediction labels
[0007] Step S4: High-confidence label filtering and updating: Set high-confidence threshold Regarding confidence level For the given samples, update the sample labels to the predicted labels: For confidence levels below For the samples, keep the original labels of the samples. constant; Step S5: low confidence sample analysis and convergence judgment: set a low confidence threshold confidence of the sample, if the prediction probability distribution of the low confidence sample is concentrated in two adjacent categories, and the probability value difference of the two adjacent categories is less than a preset threshold; it is determined that the model is in a fuzzy state for the category judgment of the corresponding sample, at the same time, the verification set accuracy is monitored, if the verification set accuracy is platformized and the low confidence sample is in a fuzzy state, it is determined that the algorithm converges; and jump to execute step S8, otherwise execute step S6; Step S6: adaptive adjustment of loss function: after each iteration or at the platform period of accuracy, increase the weight of the current violated physical rule according to the model performance, adjust the hyperparameters or rule weights of the adaptive loss function , to obtain the updated loss function ; Step S7: iteration loop: repeat steps S2 to S6 until the algorithm convergence condition in step S5 is met; Step S8: output: output the final optimized label set , model and adaptive loss function .
[0008] In some embodiments of the present disclosure, the preset threshold of the probability value of the two adjacent categories in step S5 is 0.1-0.3.
[0009] According to another aspect of the present disclosure, a geological excavability label self-generation device based on adaptive loss collaborative optimization is provided, which is suitable for a geological excavability label self-generation method based on adaptive loss collaborative optimization, comprising including a data management module for storing and scheduling original excavation time series data , current iteration label and model prediction result, the data management module adopts tensor format for batch processing; The data output end of the data management module is connected with the base learning module, which is used for receiving input data and outputting prediction results; the base learning module includes a multi-layer perception MLP structure or a one-dimensional time series convolution network 1D-CNN structure; The base learning module is connected with the adaptive loss function module, which is used for calculating the adaptive loss containing the supervised loss term and the physical law constraint term; The base learning module is connected with the label updating module, which is used for comparing the prediction confidence with the preset threshold, screening and updating the label of the high confidence sample; The base learning module is connected with the convergence judgment and loss function optimization module, and is used for monitoring the validation set accuracy and low confidence sample distribution, judging the algorithm convergence, and adjusting the hyperparameters and rule weights of the loss function. The modules work cooperatively to realize self-generation of the geological excavability label through iterative optimization.
[0010] In some embodiments of the present disclosure, the multi-layer perception (MLP) structure comprises: an input layer with a dimension of d; a first fully connected layer with 512 neurons and a ReLU activation function; a first Dropout layer with a dropout rate of 0.2; a second fully connected layer with 256 neurons and a ReLU activation function; a second Dropout layer with a dropout rate of 0.2; an output layer with C neurons, where C is the number of categories; The one-dimensional time sequence convolutional network (1D-CNN) structure comprises: an input layer with a size of d x 1; a first one-dimensional convolutional layer with 64 filters, a kernel size of 3, and a ReLU activation function; a max-pooling layer with a pooling size of 2; a second one-dimensional convolutional layer with 128 filters, a kernel size of 3, and a ReLU activation function; a global average pooling layer; an output layer with C neurons, where C is the number of categories.
[0011] In some embodiments of the present disclosure, the label updating module inputs the predicted probability distribution of all samples and the current label , and outputs the updated label . The convergence judgment and loss function optimization module is used for monitoring the training process, calculating the validation set accuracy, and analyzing the probability distribution of low confidence samples, so as to trigger the adjustment strategy of the loss function hyperparameters.
[0012] In some embodiments of the present disclosure, the adaptive loss calculated by the adaptive loss function module is represented as:
[0013] In the formula, is a supervision loss term, is a physical law constraint term, is a hyperparameter for balancing the contributions of the two terms; the supervision loss term : For C-classification problems, the cross-entropy loss function is adopted:
[0014] where, is the indicator function, is the output value of the model for the i-th sample in the j-th class; N is the number of samples Physical law constraint term By constructing a differentiable penalty term to inject engineering knowledge into the model, including the following steps: a. Map the class probability distribution of the model output to a continuous easy-to-dig exponential scalar value , so as to calculate with continuous physical quantities;
[0015] where, is the easy-to-dig value corresponding to each class; b. Based on the TBM tunneling mechanism, the following rule constraint term is constructed: b1. Thrust and speed relationship constraint term:
[0016] where, , is the normalization coefficient, is the proportion coefficient, is the thrust, is the propulsion speed; b2. Energy consumption constraint term:
[0017] where: is the specific energy; is the normalization coefficient; b3. Torque and thrust coupling constraint term:
[0018] where, is the torque, is the normalization function; is the proportion coefficient; The physical law constraint term is the weighted sum of each rule loss:
[0019] where, are all adjustable weights, In some embodiments of the present disclosure, the input data of the data management module includes time series data collected during the tunneling process of the tunnel boring machine, specifically including: Time Unix timestamp or date-time format, for data alignment and time series analysis, Ring number Integer sequence, for locating the TBM tunneling position; Total thrust kN, continuous value, reflecting the overall thrust of the cutterhead, Thrust speed , unit: mm / min, continuous value, reflecting the tunneling efficiency; Cutterhead torque , unit: kN·m, continuous value, reflecting the torque required for the cutterhead to cut the rock mass; Cutterhead speed , unit: rpm, continuous value; Penetration , unit: mm / rev, calculated from the thrust speed v and the cutterhead speed n; Shield displacement , unit: mm, reflecting the interaction between the shield and the surrounding rock; Shield pressure , unit: MPa, reflecting the interaction between the shield and the surrounding rock.
[0020] In some embodiments of the present disclosure, the data processing of the data management module includes the following steps: a. Data cleaning: using sliding window quartile range method to detect and eliminate sensor outliers.
[0021] b. Missing value processing: for missing values, using time series aware linear interpolation method to fill in: .
[0022] c. Feature engineering: c1. Construct derived features: calculate specific energy .
[0023] c2. Time series features: calculate window statistics of thrust and speed features, including moving average and standard deviation; C3. Data normalization: using Z-Score standardization to process all continuous features: , where is the mean of each feature, standard deviation.
[0024] In some embodiments of the present disclosure, the data output of the data management module includes: The optimized geological label is a vector or sequence with a length of N, representing a geological excavability index value; The trained model parameters include a weight file of the model ; The final version of the loss function includes optimized hyperparameters ; The function definition.
[0025] In some embodiments of the present disclosure, the training process of the data management module is performed through multiple iterations, and each iteration includes the following steps: a. Forward propagation: obtaining batch data and corresponding labels from the data management module; inputting into the base model to obtain output logits .
[0026] b. Loss calculation: b1. Calculate the supervised loss term
[0027] b2. Obtain the probability distribution through Softmax and map it to the continuous easy-to-dig index ESI value
[0028] b3. Extract the thrust and speed physical features from , and calculate the physical constraint term
[0029] b4. Merge the two terms to obtain the total loss
[0030] c. Backpropagation and gradient calculation: calculate the total loss with respect to the model parameters :
[0031] d. Parameter update: update the model parameters according to the calculated gradient using the optimizer: , where is the learning rate; e. Iteration loop: repeat the above steps until the model converges on the training set or reaches the preset number of iterations; f. Loss function optimization: monitor the composition of the physical loss sub-item of the model on the training set, and adjust the weight coefficient of the physical loss share in the iteration process according to the size relationship of the physical loss share.
[0032] The present application has the following advantages: 1. Originality: The framework of "model-loss-label" co-evolution is proposed, which optimizes model parameters, loss function and output label simultaneously through iterative method, solving the learning problem under unlabeled or weak labeled industrial data.
[0033] 2. High precision and high reliability: Through physical constraint loss and high confidence screening mechanism, the physical reasonableness and statistical reliability of the generated label are double guaranteed, and its precision is much better than the initial label.
[0034] 3. Strong generalization: The adaptive loss function mechanism enables the algorithm to automatically adjust the learning goal according to the characteristics of different engineering data, avoiding model overfitting to biased initial labels, and having stronger generalization ability.
[0035] 4. Strong engineering practicability: The algorithm input is all TBM routine monitoring data, without additional sensors, and the output result can be directly used to guide TBM intelligent tunneling parameter setting, with strong landing performance.
[0036] The self-generation of high-quality geological labels is realized, and through the iterative optimization mechanism, accurate and reliable geological drillability labels can be automatically generated from unlabeled TBM tunneling data, effectively solving the problems of lack of label data, high cost of manual labeling and strong subjectivity in this field. The generalization ability and interpretability of the model are improved, the relationship between thrust and speed, energy consumption, torque-thrust coupling and other field knowledge are embedded in the loss function in the form of differentiable physical law constraint terms, guiding the model to learn the feature representation conforming to the engineering mechanism, so that the prediction result not only has high precision, but also has clear physical meaning, enhancing the reliability and credibility of the model. A closed loop of label, model and loss function co-optimization is formed, and the label updating, model training and loss function adjustment are creatively placed in a unified iterative framework. High confidence prediction is used for iterative optimization of labels, and low confidence samples are used to guide loss function adjustment, while the optimized loss function further improves the model performance. The three co-evolve, and finally converge to the optimal solution. It has good engineering applicability and automation degree, provides a complete automation process from data processing, model training to convergence judgment, reduces the need for manual intervention, and provides an online deployable intelligent solution for TBM intelligent tunneling. BRIEF DESCRIPTION OF DRAWINGS
[0037] Figure 1 The flow chart of the self-generation method of geological drillability label based on adaptive loss co-optimization; DETAILED DESCRIPTION
[0038] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and not to limit the present application. Example 1
[0039] This example discloses a self-generation method for geological excavability labels based on adaptive loss collaborative optimization. See [link to relevant documentation]. Figure 1 The steps include: Step S1: Data preparation and initialization: Input an unlabeled TBM tunneling time series dataset. The D contains N samples, denoted as In the formula Given a d-dimensional feature vector, initial geological labels are generated for the dataset D through random assignment or by utilizing a weak engineering model. In the formula Provide discrete level labels or continuous value labels; initialize the adaptive loss function. Base learning model ; Step S2: Model Training: In the first step... In the next iteration, the current dataset is used. Current loss function The base learning model in training step S1 The training objective is to minimize the loss function. In the formula For the model parameter; Step S3: Prediction and Confidence Generation: Using the trained model For dataset Make predictions and obtain the prediction output for each sample. The probability distribution for each category, through Calculate the confidence score for each sample. and prediction labels
[0040] Step S4: High-confidence label filtering and updating: Set high-confidence threshold Regarding confidence level For the given samples, update the sample labels to the predicted labels: For confidence levels below For the samples, keep the original labels of the samples. constant; Step S5: Low-confidence sample analysis and convergence judgment: Set a low-confidence threshold Identify confidence If the predicted probability distribution of low-confidence samples is concentrated in two adjacent categories, and the difference in probability values between the two adjacent categories is less than a preset threshold, then the model is determined to be in a state of ambiguity in class identification of the corresponding samples. At the same time, the accuracy of the validation set is monitored. If the accuracy of the validation set is plateaued and the low-confidence samples are in a state of ambiguity, then the algorithm is determined to be converged and the process jumps to step S8; otherwise, step S6 is executed. Step S6: loss function adaptive adjustment: at the end of each iteration or accuracy plateau, increase the weight of the current violated physical rule according to the model performance, adjust the adaptive loss function of the hyperparameters or rule weights to obtain an updated loss function ; Step S7: iterative loop: repeat steps S2 to S6 until the algorithm converges in step S5 Step S8: output: output the final optimized label set , model and adaptive loss function .
[0041] The preset threshold of the probability value of the two adjacent categories in step S5 is 0.1-0.3 According to another aspect of the present disclosure, a geological excavability label self-generation device based on adaptive loss collaborative optimization is provided, which is suitable for a geological excavability label self-generation method based on adaptive loss collaborative optimization, comprising including a data management module for storing and scheduling original excavation time series data , the label of the current iteration and the model prediction result, the data management module adopts tensor format for batch processing; The data output end of the data management module is connected with the base learning module, which is used for receiving input data and outputting prediction results; the base learning module includes a multi-layer perception MLP structure or a one-dimensional time series convolution network 1D-CNN structure; The base learning module is connected with the adaptive loss function module, which is used for calculating the adaptive loss containing the supervised loss term and the physical law constraint term; The base learning module is connected with the label updating module, which is used for comparing the prediction confidence with the preset threshold, filtering and updating the label of the high confidence sample; The base learning module is connected with the convergence judgment and loss function optimization module, which is used for monitoring the validation set accuracy and low confidence sample distribution, judging the algorithm convergence, and adjusting the hyperparameters and rule weights of the loss function; Each module works cooperatively to realize the self-generation of the geological excavability label through iterative optimization.
[0042] The multi-layer perception MLP structure comprises: input layer, dimension d; The first full connection layer has 512 neurons, and the activation function is ReLU; The first Dropout layer has a dropout rate of 0.2; The second full connection layer has 256 neurons, and the activation function is ReLU; The second Dropout layer has a dropout rate of 0.2; The output layer has C neurons, and C is the number of categories; The one-dimensional time sequence convolutional network 1D-CNN structure comprises: The input layer has a size of d x 1; The first one-dimensional convolutional layer has 64 filters, a kernel size of 3, and a ReLU activation function; The maximum pooling layer has a pooling size of 2; The second one-dimensional convolutional layer has 128 filters, a kernel size of 3, and a ReLU activation function; The global average pooling layer; The output layer has C neurons, and C is the number of categories.
[0043] The label updating module inputs the predicted probability distribution of all samples and the current label , and outputs the updated label ; The convergence judgment and loss function optimization module is used to monitor the training process, calculate the validation set accuracy, and analyze the probability distribution of low confidence samples to trigger the adjustment strategy of the loss function hyperparameters.
[0044] The adaptive loss calculated by the adaptive loss function module is represented as:
[0045] In the formula, is a supervision loss term, is a physical law constraint term, is a hyperparameter supervision loss term that balances the contributions of the two terms: For a C-classification problem, a cross-entropy loss function is used:
[0046] In the formula, is an indicator function, is the output value of the model for the th sample in the th category; N is the number of samples The physical law constraint term injects engineering knowledge into the model by constructing a differentiable penalty term, including the following steps: a. Map the category probability distribution output by the model to a continuous, easy-to-dig exponential scalar value so as to be calculated with continuous physical quantities;
[0047] wherein, is the easy-to-dig value corresponding to each category; b. Based on the TBM tunneling mechanism, the following rule constraint terms are constructed: b1. Thrust and speed relationship constraint term:
[0048] wherein, , is a normalization coefficient, is a proportional coefficient, is the thrust, is the advancing speed; b2. Energy consumption constraint term:
[0049] wherein: is the specific energy; is a normalization coefficient; b3. Torque and thrust coupling constraint term:
[0050] wherein, is the torque, is a normalization function; is a proportional coefficient; the physical law constraint term is the weighted sum of each rule loss:
[0051] wherein, are all adjustable weights, The input data of the data management module includes the time series data collected during the tunneling process of the tunneling machine, specifically including: Time Unix timestamp or date and time format, used for data alignment and time series analysis, Ring number Integer sequence, used to locate the TBM tunneling position; Total thrust kN, continuous value, reflecting the overall advancing force of the cutterhead, Advancing speed , unit: mm / min, continuous value, reflecting the tunneling efficiency; Cutterhead torque , unit: kN·m, continuous value, reflecting the torque required for the cutterhead to cut the rock mass; Cutterhead rotational speed , unit: rpm, continuous value; Penetration , unit: mm / rev, calculated from the thrusting speed v and the cutterhead rotation speed n; Shield displacement , unit: mm, reflecting the interaction between the shield and the surrounding rock; Shield pressure , unit: MPa, reflecting the interaction between the shield and the surrounding rock.
[0052] The data processing of the data management module includes the following steps: a. Data cleaning: using the sliding window quartile range method to detect and eliminate sensor outliers.
[0053] b. Missing value processing: for missing values, using the linear interpolation method of time series perception to fill in: .
[0054] c. Feature engineering: c1. Construct derived features: calculate specific energy .
[0055] c2. Time series features: calculate the window statistics of thrust and speed features, including moving average and standard deviation; C3. Data normalization: use Z-Score standardization to process all continuous features: , where is the mean of each feature, standard deviation.
[0056] The data output of the data management module includes: The optimized geological label is a vector or sequence of length N, representing the geological excavability index value. The trained model parameters include the weight file of the model ; The final version of the loss function includes the function definition of the optimized hyperparameters .
[0057] The training process of the data management module is performed through multiple iterations, and each iteration includes the following steps: a. Forward propagation: obtain batch data and corresponding labels from the data management module; input into the base model , get the output logits .
[0058] b. Loss calculation: b1. Calculate the supervised loss term
[0059] b2. Calculate the Get probability distribution by Softmax, and map to continuous easy-digging index ESI value
[0060] b3. Extract thrust and speed physical features from , calculate physical constraint term
[0061] b4. Merge two terms to get total loss
[0062] c. Backpropagation and gradient calculation: calculate total loss Gradient of model parameters :
[0063] d. Parameter update: update model parameters according to calculated gradient using optimizer: , where is learning rate; e. Iteration loop: repeat the above steps until the model converges on the training set or reaches the preset number of iterations; f. Loss function optimization: monitor the composition of the physical loss sub-item of the model on the training set, and adjust its weight coefficient in the iteration process according to the size relationship of the physical loss share.
[0064] Although some preferred embodiments of the present application have been described, those skilled in the art, once they know the basic creative concept, can make additional changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.
[0065] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.
Claims
1. A geological excavatability label self-generation method based on adaptive loss collaborative optimization, characterized in that, The method comprises the following steps: Step S1 : Data preparation and initialization: input unlabeled TBM drilling timing dataset , the D contains N samples, denoted as , in which is a d-dimensional feature vector, and initial geological labels are generated for the dataset D by random assignment or using a weak engineering model , in which is a discrete level label or a continuous value label; initialize the adaptive loss function and the base learning model ; Step S2: Model training: In the first iteration, the base learning model in step S1 is trained using the current dataset current loss function ; The training objective is to minimize the loss function: where is the model parameters; Step S3: Prediction with confidence generation: Using the trained model On the dataset Make predictions, getting the predicted output for each sample For each class probability distribution, by Compute the confidence for each sample And the predicted label ; Step S4: High confidence label filtering and updating: set a high confidence threshold ; for samples with confidence , update the sample label to the predicted label: For samples with confidence lower than , keep the original label of the sample unchanged Step S5: low confidence sample analysis and convergence judgment: set a low confidence threshold Recognition confidence If the prediction probability distribution of the low confidence sample is concentrated in two adjacent categories, and the probability value difference between the two adjacent categories is less than a preset threshold, it is determined that the category judgment of the corresponding sample by the model is in a fuzzy state. At the same time, the verification set accuracy is monitored. If the verification set accuracy is flat and the low confidence sample is in a fuzzy state, it is determined that the algorithm converges; and jump to execute step S8, otherwise execute step S6; Step S6: Loss function self-adaptation adjustment: at the end of each iteration or at the accuracy plateau, adjust the adaptive loss function according to the model performance to increase the weight of the current violated physical rule, obtain the updated loss function of the hyperparameters or rule weights ; Step S7: Iterative loop: repeat steps S2 to S6 until the algorithm convergence condition in step S5 is met; Step S8: Output: Output the final optimized label set , model and adaptive loss function .
2. The method of claim 1, wherein the method is based on adaptive loss cooperative optimization. The preset threshold of the probability values of two adjacent categories in step S5 is 0.1-0.
3.
3. A geological excavatability label self-generation device based on adaptive loss collaborative optimization, applicable to the geological excavatability label self-generation method based on adaptive loss collaborative optimization as claimed in claim 1, characterized in that: The data management module is configured to store and schedule raw excavation timing data a label of the current iteration and a model prediction result, the data management module is configured to batch process in a tensor format; The data output end of the data management module is connected to the base learning module, which is used to receive input data and output prediction results; the base learning module comprises a multi-layer perception machine (MLP) structure or a one-dimensional time series convolutional network (1D-CNN) structure; The base learning module is connected to the adaptive loss function module, which is used to calculate the adaptive loss containing a supervised loss term and a physical law constraint term; The base learning module is connected to the label update module, which is used to compare the prediction confidence with a preset threshold, filter and update the labels of high-confidence samples; The base learning module is connected to the convergence judgment and loss function optimization module, which is used to monitor the validation set accuracy and low-confidence sample distribution, judge the algorithm convergence, and adjust the hyperparameters and rule weights of the loss function; Each module works collaboratively to realize the self-generation of the geological excavatability label through iterative optimization.
4. The device for self-generating geological excavatability labels based on adaptive loss collaborative optimization according to claim 3, wherein: The multi-layer perception machine (MLP) structure comprises: an input layer with a dimension of d; a first fully connected layer with 512 neurons and a ReLU activation function; a first Dropout layer with a dropout rate of 0.2; a second fully connected layer with 256 neurons and a ReLU activation function; a second Dropout layer with a dropout rate of 0.2; an output layer with C neurons, where C is the number of categories; The one-dimensional time series convolutional network (1D-CNN) structure comprises: an input layer with a size of d x 1; a first one-dimensional convolutional layer with 64 filters, a kernel size of 3, and a ReLU activation function; a max pooling layer with a pooling size of 2; a second one-dimensional convolutional layer with 128 filters, a kernel size of 3, and a ReLU activation function; a global average pooling layer; an output layer with C neurons, where C is the number of categories.
5. The device of adaptive loss-cooperative optimization based geological excavatability label self-generation of claim 3, wherein: The label updating module inputs the predicted probability distribution for all samples and the current label , and outputs the updated label ; The convergence judgment and loss function optimization module is used to monitor the training process, calculate the validation set accuracy, and analyze the probability distribution of low-confidence samples, triggering the adjustment strategy of the loss function hyperparameters.
6. The self-adaptive loss collaborative optimization based geobility label self-generation device of claim 3, wherein: The adaptive loss function module calculates an adaptive loss is represented as: ； wherein, is a supervision loss term, is a physical law constraint term, is a hyperparameter balancing the contributions of both terms; Supervised loss term : For a C-classification problem, a cross-entropy loss function is used: ； wherein is an indicator function, is the output value of the model for the th sample in the th class; N is the number of samples Physical law constraint term Engineering knowledge is injected into the model by constructing a differentiable penalty term, including the following steps: a. mapping the class probability distribution output by the model to a continuous ease-of- excavation index scalar value so as to be calculated with continuous physical quantities; ； In the formula, is the value of the easy-to-dig property corresponding to each category; b. Based on the TBM excavation mechanism, the following rule constraint terms are constructed: b1. Thrust and speed relationship constraint term: ； wherein, , is a normalization coefficient, is a proportionality coefficient, is a thrust force, is a propulsion speed; b2. Energy consumption constraint term: ； In the formula: is the specific energy; is the normalization coefficient; b3. Torque and thrust coupling constraint term: ； wherein is the torque, is a normalization function; is a proportionality coefficient; the physical law constraint term is a weighted sum of the individual rule losses: ； wherein, are all adjustable weights.
7. The self-adaptive loss collaborative optimization based geobearability label self-generation device of claim 3, wherein: The input data of the data management module includes time series data collected during the tunneling process of the tunneling machine, specifically including: Time Unix timestamp or date-time format, for data alignment and timing analysis, Ring number Integer sequence for positioning the TBM drilling location; Total thrust kN, is a continuous value reflecting the overall thrust of the cutterhead, Advancing speed in mm / min, continuous value, reflects the tunneling efficiency; Torque of cutterhead , unit: kN.m, continuous value, reflecting the torque required for the cutterhead to cut rock mass; Cutterhead rotational speed in rpm, as a continuous value; penetration in mm / rev, calculated from the advance speed v and the disc rotational speed n; Shield displacement in mm, reflecting the interaction of the shield with the surrounding rock; Shield pressure in MPa, reflecting the interaction between the shield and the surrounding rock.
8. The self-adaptive loss collaborative optimization based geobearability label self-generation device of claim 3, wherein: The data processing of the data management module comprises the following steps: a. Data cleaning: use the sliding window quartile range method to detect and eliminate sensor outliers; b. Dealing with missing values: For missing values, linear interpolation with time awareness is used to fill in: ; c. Feature engineering: c1. Construct derived features: compute specific energy ; c2. Time series features: calculate the window statistics of the thrust and speed features, including the moving average and standard deviation; C3. Data Normalization: Z-Score standardization is applied to all continuous features: where is the mean, is the standard deviation for each feature.
9. The self-adaptive loss cooperative optimization based geobility label self- generation device of claim 3, wherein: The data output of the data management module includes: The optimized geological label is a vector or sequence with a length of N, representing a geological excavability index value; trained model parameters, including a weight file of the model Final loss function, including optimized hyperparameters of Function definition.
10. The self-adaptive loss collaborative optimization based geobility label self- generation device of claim 3, wherein: The training process of the data management module is performed through multiple iterations, and each iteration includes the following steps: a. Forward pass: get batch data from data management module and corresponding labels ; input to base model , get output logits ; b. Loss calculation: b1. computing a supervision loss term ; b2. will be Get probability distribution by Softmax, and map to continuous easy-to-dig index ESI value ; b3. From extracting the thrust and speed physical characteristics, calculating the physical constraint term ; b4. Merge both, get total loss ; c. Backpropagation and gradient computation: Compute total loss with respect to model parameters ; d. Parameter update: Update the model parameters using the optimizer according to the computed gradients: where is the learning rate; e. Iteration loop: repeat the above steps until the model converges on the training set or reaches a preset number of iterations; f. Loss function optimization: Monitor the physical loss components of the model on the training set The composition of the physical loss is adjusted in the iteration process according to the size relationship of the sharing of the physical loss.