A filling body stage intelligent identification method and system based on multi-model integration and transfer learning

Abstract

The present application belongs to the technical field of cross of mine safety monitoring, rock mechanics and artificial intelligence, and particularly relates to a filling body stage intelligent identification method and system based on multi-model integration and transfer learning. First, the acoustic emission monitoring system of the filling body is built to collect the acoustic emission signals of the filling body in the whole loading process and extract multi-dimensional features; the improved differential evolution algorithm is used to optimize the piecewise linear regression model, and the four stages of filling body failure are automatically identified in combination with physical constraints to generate the source domain data set; then the key features are extracted through the anti-leakage feature optimization mechanism, and a soft voting heterogeneous ensemble model is constructed by using four kinds of heterogeneous base learners; aiming at the difference of data distribution of the target domain, the domain offset degree is calculated and the fine-tuning transfer strategy based on instance-level weighting is adopted to quickly update the model with a small amount of data. The present application solves the problems of strong subjectivity of artificial labeling, poor generalization of single model and difficult cross-domain adaptation, and realizes intelligent early warning of the whole process of filling body instability.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of mine safety monitoring, rock mechanics and artificial intelligence. Specifically, it relates to an intelligent identification method and system for the stage of backfilling based on multi-model integration and transfer learning. It utilizes acoustic emission technology combined with multi-model integration learning algorithms and transfer learning strategies to perform fully automatic stage identification and instability early warning of the loading and failure process of mine backfilling. Background Technology

[0002] With the depletion of shallow mineral resources, deep mining has become an inevitable trend in global mining development. Backfill mining is widely used due to its ability to effectively control ground pressure and reduce surface subsidence. As the core structure supporting the goaf, the stability of the backfill directly affects the safe production of the mine. During the loading process of the backfill, the initiation, propagation, and eventual macroscopic failure of internal microcracks release transient elastic waves, i.e., acoustic emission (AE) signals. By monitoring these signals, the damage evolution state of the backfill can be inverted. However, existing acoustic emission monitoring and stage identification technologies face many challenges, mainly manifested in low data annotation efficiency, strong annotation subjectivity, insufficient model generalization ability, and severe domain offset problems.

[0003] First, data annotation suffers from inefficiency and high subjectivity. Traditional supervised learning models rely on experts manually annotating damage stages in acoustic emission signals, and the annotation results from different experts vary, affecting the model's training effectiveness. Second, single machine learning models struggle to handle the complexity of acoustic emission signals from infill bodies. Existing research often employs models such as Support Vector Machines (SVM) and Backpropagation Neural Networks (BPNN), but these models lack generalization ability when faced with complex nonlinear features and noise interference, easily leading to overfitting or underfitting. Finally, the domain shift problem makes it difficult to directly transfer models trained on standard laboratory data to target monitoring environments. Since the data distribution in the target domain often differs significantly from that in the source domain, traditional models cannot adapt to this cross-domain change, resulting in a sharp decline in model performance or even model failure in real-world applications.

[0004] Therefore, there is an urgent need for an automated and efficient multi-model integration method that can solve problems such as inconsistent annotations, insufficient generalization ability of single models, and domain offset, thereby improving the accuracy and robustness of intelligent recognition in the filling stage. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides an intelligent identification method and system for the filling stage based on multi-model integration and transfer learning. It aims to solve the problem of manual annotation by constructing an intelligent monitoring system for the entire process, utilizing the global optimization capability of differential evolution algorithm to solve the problem of manual annotation, improving the identification accuracy through multi-model integration strategy, and solving the cross-domain adaptation problem through weighted fine-tuning transfer technology.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for intelligent identification of infill body stages based on multi-model ensemble and transfer learning, the method comprising: The constructed acoustic emission monitoring system for infill bodies was used to collect acoustic emission signals from the source domain infill bodies and perform noise reduction processing. Based on the noise-reduced acoustic emission signals, multi-dimensional feature vectors were extracted to construct the original feature dataset. An adaptive labeling model based on differential evolution algorithm is constructed to automatically identify three critical time breakpoints in the four stages of the source domain filling failure process, generate stage labels for the source domain filling failure process, and obtain the source domain dataset. Based on the source domain dataset, a leakage prevention feature optimization mechanism is used to filter the features in the original feature dataset. The filtered features are then used to construct a soft voting heterogeneous ensemble model by fusing Bagging and Boosting algorithms, and the source domain fault diagnosis model is trained. Acoustic emission signals of the target domain filling body are collected to obtain the target domain dataset. The maximum mean difference between the target domain dataset and the source domain dataset is calculated. If the maximum mean difference exceeds a preset threshold or the test accuracy of the source domain fault diagnosis model on the target domain dataset is lower than the preset baseline value, a significant domain shift is determined and the transfer learning process is triggered. A hybrid dataset is constructed based on the target domain dataset and the source domain dataset. Based on the hybrid dataset, a weighted fine-tuning migration strategy is used to update the source domain fault diagnosis model to obtain the final infill body instability stage identification model. The infill body instability stage identification model is used to realize intelligent identification of the infill body stage.

[0007] Preferably, the multidimensional feature vector includes: Basic parameter characteristics: ring count, energy, amplitude, duration, rise time; Frequency domain characteristics: center frequency, peak frequency, average frequency, and frequency centroid; Waveform characteristics: RA value, AF value; Evolutionary characteristics: the cumulative value of basic parameters over time, the rate of change, and the event density.

[0008] Preferably, the method for constructing an adaptive annotation model based on differential evolution algorithm includes: ; In the formula, Let be the objective function. There are three critical time breakpoints. For the first k Sum of squared fitting residuals within each stage The penalty coefficient is... P This is a physical constraint penalty term.

[0009] Preferably, the method for calculating the maximum mean difference between the target domain dataset and the source domain dataset includes: ; In the formula, For the maximum mean difference, and These represent the source domain dataset and the target domain dataset, respectively. and These represent the number of samples in the source and target domains, respectively. and The source domain is respectively The nth sample and the target domain The feature vector of each sample Let represent the mapping function to the reproducing kernel Hilbert space. denoted as the norm in the regenerated kernel Hilbert space.

[0010] The present invention also provides an intelligent recognition system for filling body stages based on multi-model integration and transfer learning. The system is used to implement the aforementioned method and includes: a data acquisition and extraction module, a stage labeling module, a screening and construction module, an offset detection module, and a model update module. The acquisition and extraction module is used to acquire the acoustic emission signal of the source domain filling body using the constructed filling body acoustic emission monitoring system and perform noise reduction processing. Based on the noise-reduced acoustic emission signal, multi-dimensional feature vectors are extracted to construct the original feature dataset. The stage labeling module is used to build an adaptive labeling model based on the differential evolution algorithm, automatically identify the three critical time breakpoints of the four stages of the source domain filling failure process, generate stage labels for the source domain filling failure process, and obtain the source domain dataset. The filtering module is used to filter features in the original feature dataset based on the source domain dataset using a leak-proof feature optimization mechanism. The filtered features are then used to construct a soft-voting heterogeneous ensemble model by fusing Bagging and Boosting algorithms, and the source domain fault diagnosis model is trained. The offset detection module is used to collect the acoustic emission signal of the target domain filling body to obtain the target domain dataset, calculate the maximum mean difference between the target domain dataset and the source domain dataset, and if the maximum mean difference exceeds the preset threshold or the test accuracy of the source domain fault diagnosis model on the target domain dataset is lower than the preset baseline value, it is determined that there is a significant domain offset and the transfer learning process is triggered. The model update module is used to construct a hybrid dataset based on the target domain dataset and the source domain dataset. Based on the hybrid dataset, a weighted fine-tuning migration strategy is used to update the source domain fault diagnosis model to obtain the final infill body instability stage identification model. The infill body instability stage identification model is used to realize intelligent identification of the infill body stage.

[0011] Preferably, the multidimensional feature vector includes: Basic parameter characteristics: ring count, energy, amplitude, duration, rise time; Frequency domain characteristics: center frequency, peak frequency, average frequency, and frequency centroid; Waveform characteristics: RA value, AF value; Evolutionary characteristics: the cumulative value of basic parameters over time, the rate of change, and the event density.

[0012] Preferably, the method for constructing an adaptive annotation model based on differential evolution algorithm includes: ; In the formula, Let be the objective function. There are three critical time breakpoints. For the first k Sum of squared fitting residuals within each stage The penalty coefficient is... P This is a physical constraint penalty term.

[0013] Preferably, the method for calculating the maximum mean difference between the target domain dataset and the source domain dataset includes: ; In the formula, For the maximum mean difference, and These represent the source domain dataset and the target domain dataset, respectively. and These represent the number of samples in the source and target domains, respectively. and The source domain is respectively The nth sample and the target domain The feature vector of each sample Let represent the mapping function to the reproducing kernel Hilbert space. denoted as the norm in the regenerated kernel Hilbert space.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Achieves fully automated and objective data annotation: The adaptive annotation algorithm based on differential evolution proposed in this invention utilizes the physical evolution law of acoustic emission cumulative energy to transform stage division into a mathematical optimization problem. Compared with manual annotation, this method eliminates subjective errors, has a fast processing speed, and can ensure a high degree of consistency in the annotation standards of all samples, providing high-quality "benchmark truth values" for subsequent model training.

[0015] 2. Significantly improves the model's generalization ability and noise robustness: By integrating four heterogeneous models—RandomForest, ExtraTrees, XGBoost, and LightGBM—this invention fully combines the advantages of the Bagging algorithm in reducing variance and the Boosting algorithm in reducing bias. The soft voting mechanism enables the ensemble model to make more robust decisions than a single model when facing complex nonlinear acoustic emission characteristics, effectively avoiding overfitting.

[0016] 3. The weighted fine-tuning migration strategy proposed in this invention innovatively solves the domain offset problem. By introducing the maximum mean difference as a offset metric and using a very small amount of target domain data to perform targeted updates on the ensemble model, the ensemble model can quickly adapt to the specific distribution of the target domain while retaining the general physical laws learned from the source domain. This greatly reduces the cost and time required in practical applications.

[0017] 4. Comprehensive and efficient feature engineering system: A multi-dimensional feature space containing 35 features was constructed, and a leakage prevention feature selection mechanism was introduced. Experiments show that the selected key features such as event density and duration have clear physical meaning, effectively improving computational efficiency and recognition accuracy.

[0018] In summary, this invention provides an innovative intelligent identification method for infill stage based on multi-model integration and transfer learning, which can monitor and identify the failure stage of infill structure in real time and accurately, and has important theoretical significance and practical application value. Attached Figure Description

[0019] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the overall process of the method in this embodiment of the invention, which details the entire logic from data acquisition, automatic annotation, source domain modeling to transfer learning. Figure 2 This is a schematic diagram of the adaptive stage division principle based on the differential evolution algorithm in an embodiment of the present invention, showing the fitting effect of the cumulative energy curve and the automatic segmentation results of the four stages; Figure 3 The image shows the feature importance ranking and filtering results of an embodiment of the present invention, displaying the Top-18 key features and their importance scores; Figure 4The confusion matrix diagram of the source domain fault diagnosis model in this embodiment of the invention on the test set shows the model's recognition accuracy for the four stages; Figure 5 This is a comparison chart of the accuracy of the transfer learning strategy with the zero-sample baseline under different label ratios in the embodiments of the present invention, which verifies the effectiveness of transfer learning. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Example 1 like Figure 1 As shown, this invention provides an intelligent recognition method for the filling stage based on multi-model ensemble and transfer learning, comprising: The constructed acoustic emission monitoring system for infill bodies was used to collect acoustic emission signals from the source domain infill bodies and perform noise reduction processing. Based on the noise-reduced acoustic emission signals, multi-dimensional feature vectors were extracted to construct the original feature dataset. An adaptive labeling model based on differential evolution algorithm is constructed to automatically identify three critical time breakpoints in the four stages of the source domain filling failure process, generate stage labels for the source domain filling failure process, and obtain the source domain dataset. Based on the source domain dataset, a leakage prevention feature optimization mechanism is used to filter the features in the original feature dataset. The filtered features are then used to construct a soft voting heterogeneous ensemble model by fusing Bagging and Boosting algorithms, and the source domain fault diagnosis model is trained. Acoustic emission signals of the target domain filling body are collected to obtain the target domain dataset. The maximum mean difference between the target domain dataset and the source domain dataset is calculated. If the maximum mean difference exceeds a preset threshold or the test accuracy of the source domain fault diagnosis model on the target domain dataset is lower than the preset baseline value, a significant domain shift is determined and the transfer learning process is triggered. A hybrid dataset is constructed based on the target domain dataset and the source domain dataset. Based on the hybrid dataset, a weighted fine-tuning migration strategy is used to update the source domain fault diagnosis model to obtain the final infill body instability stage identification model. The infill body instability stage identification model is used to realize intelligent identification of the infill body stage.

[0024] The specific implementation process of this invention is as follows: In practical implementation, real-time data of the filling material is collected and intelligently analyzed using a filling material acoustic emission monitoring system. This invention embodiment prepared two sets of filling material samples with significantly different physical and mechanical properties for experiments: The first mix proportion filling sample (source area filling): PO 42.5 grade ordinary Portland cement and all tailings were used, with a cement-to-sand ratio of 1:4 and a slurry mass concentration of 75%. Standard cubic test blocks, at temperature Cured for 28 days in a standard environment with humidity above 90%. Its average uniaxial compressive strength is 3.85 MPa.

[0025] The second mix proportion filling sample (target domain filling): using the same raw materials, with a cement-to-sand ratio of 1:8 and a slurry mass concentration of 72%, and cured for 28 days under the same conditions. Its average uniaxial compressive strength was 1.05 MPa. Due to the reduced cement content and increased moisture content, its acoustic emission signal showed significant differences from the first mix proportion in amplitude, frequency, and attenuation evolution, simulating the model failure scenario caused by fluctuations in the mine filling process.

[0026] Specific methods include: S1. The constructed acoustic emission monitoring system for the filling material is used to collect acoustic emission signals of the source domain filling material during the entire uniaxial compression loading process and perform noise reduction processing. Based on the noise-reduced acoustic emission signals, multi-dimensional feature vectors, including time domain, frequency domain, and waveform parameters, are extracted to construct the original feature dataset. Specifically, this includes: A high-sensitivity acoustic emission monitoring system for the filling material was constructed: an R15a narrowband sensor (center frequency 150kHz) was selected and coupled to the surface of the filling material in the source region. The preamplifier gain was set to 40dB, and the threshold value was set to 45dB to filter out ambient background noise. During uniaxial compression, the sensor received elastic wave signals generated by the initiation and propagation of microcracks inside the filling material in the source region.

[0027] Extracting basic parameter features: For each acoustic emission event exceeding a threshold value (denoted as the th), i (Early sound emission events), calculate ring count. (Characterizing the frequency of crack activity) and energy (Characterizing the elastic energy released by the crack): In the formula, and The first The start and end times of an acoustic emission event. The voltage signal acquired by the sensor. t It is a time variable.

[0028] Waveform features reflecting crack properties, namely RA and AF values, are extracted to distinguish the failure modes of the source region infill. In the formula, For the rising time, For the maximum amplitude, For duration. Physically, a high RA value / low AF value corresponds to shear cracks caused by aggregate displacement within the source domain infill, while a low RA value / high AF value corresponds to tensile cracks caused by matrix cracking.

[0029] Construct a 35-dimensional feature vector containing time-domain, frequency-domain, and evolutionary features. The original feature dataset: Basic parameter characteristics: ring count, energy, amplitude, duration, rise time, signal strength; Frequency domain characteristics: center frequency, peak frequency, average frequency, frequency centroid, frequency standard deviation; Waveform characteristics: RA value, AF value; Evolutionary characteristics: the cumulative value, rate of change, statistical moments, and composite ratio of basic parameters over time.

[0030] To comprehensively capture the dynamic trend of the infill body transitioning from a steady state to an unsteady state, based on the aforementioned fundamental parameter characteristics, this embodiment constructs a multi-level evolution index system: First, it calculates the cumulative and rate characteristics of evolution over time, including cumulative energy, cumulative ring count, ring count rate, event density, and energy-time density; second, it calculates the statistical characteristics of the fundamental parameters within a unit time window, including average ring count, maximum ring count, count standard deviation, maximum energy, minimum energy, and energy range; finally, it calculates the composite ratio characteristics between multiple physical quantities, including the energy-amplitude ratio (the ratio of energy to amplitude) and the count-energy ratio (the ratio of ring count to energy).

[0031] S2. Construct an adaptive labeling model based on the differential evolution algorithm to automatically identify three critical time breakpoints in the four stages of the source domain filling failure process, generate stage labels for the source domain filling failure process, and obtain the source domain dataset. Specifically, this includes: The adaptive stage division principle based on differential evolution algorithm is as follows: Figure 2 As shown, in specific implementation, the energy of a single acoustic emission event of the source domain filling body extracted in step S1 is used. Based on this, a cumulative energy curve of acoustic emission over the entire time domain is constructed. As input to the model.

[0032] Defined as cutoff time The total energy of all events that have occurred can be expressed mathematically as follows: In the formula, Indicates the cutoff time. The total number of acoustic emission events that have occurred.

[0033] Based on the physical mechanism of uniaxial compression failure of filling materials, it is assumed that the entire process includes four stages: compaction, elasticity, damage, and failure, and that the cumulative energy growth within each stage approximately follows a linear law, but the slope (energy release rate) differs significantly. To accurately identify the boundaries of these four stages, three critical time breakpoints are defined. The entire process is divided into four consecutive intervals, and an objective function containing physical constraints is constructed. To measure the piecewise fitting effect.

[0034] The objective function It is composed of the weighted sum of squared residuals from the piecewise linear regression model and the physical constraint penalty term, and its mathematical expression is: In the formula, Indicates the first The sum of squared residuals within each stage is used to quantify the deviation between the linear model and the actual cumulative energy data. Minimizing this term allows the model to approximate the actual damage evolution trajectory to the greatest extent. The penalty coefficient is set to its maximum value to force the removal of unreasonable solutions. This is a physical constraint penalty, used to limit the duration of any stage to no less than the total duration. 5%. If the combination of breakpoints found results in any stage being too short, it is determined to violate the laws of physics, and... To make the objective function value approach infinity; conversely, to make it approach infinity. .

[0035] Given the objective function Since the problem is non-convex and discontinuous across the entire time domain, this invention employs the Differential Evolutionary Algorithm (DE) in the continuous full-time domain. Internal three critical time breakpoints Perform an iterative search. The specific parameter settings and process of the DE algorithm are as follows: Population initialization: Set the population size to 50, and randomly generate 50 potential time breakpoint combinations in the solution space; Mutation operation: Set the mutation factor to 0.5, perform differential mutation on the individuals in the population to generate mutation vectors to explore the diversity of the solution space; Crossover operation: Set the crossover probability to 0.7, recombine the mutation vectors with the original individual vectors to generate experimental vectors; Selection and iteration: Compare the objective functions of the experimental vectors with those of the original individuals. , retain Smaller individuals enter the next generation. This process continues in a loop until the objective function is achieved. It converges to the minimum value or reaches the preset number of iterations.

[0036] Output the three optimal critical time breakpoints. These three breakpoints physically correspond to the transition point from pore compaction to elastic deformation of the filling material, the point of sudden increase in energy release rate caused by microcrack initiation, and the instability initiation point before macrocrack penetration. Based on these three breakpoints, the source domain data samples are sequentially divided into four stages: compaction, elasticity, damage, and failure, and assigned corresponding labels, thereby obtaining an automatically labeled source domain dataset.

[0037] S3. Based on the source domain dataset, a leakage prevention feature selection mechanism is used to filter features in the original feature dataset. The filtered features are then fused using Bagging and Boosting algorithms to construct a soft-voting heterogeneous ensemble model, which is then trained to obtain the source domain fault diagnosis model. Specifically, this includes: The source domain dataset is divided into a source domain training set and a source domain test set.

[0038] To ensure the objectivity of feature selection and prevent data leakage, this invention employs a "leakage-proof feature selection mechanism": an ExtraTreesClassifier model is trained using only the source domain training set data, and the Gini importance of each feature in the original feature dataset is calculated as the feature importance score. Figure 3 As shown, the results indicate that "duration" and "event density" are the most important, which is consistent with the high-frequency, long-duration signal characteristics caused by microcrack penetration when the source domain infill is near failure. A feature subset is constructed by selecting the Top-18 key features based on the score ranking. In practice, this feature subset is applied to the source domain test set and the subsequent target domain dataset. Specifically, this involves forced alignment of the feature space: based on the names and indices of the Top-18 key features, corresponding feature columns are extracted from the source domain test set and the target domain dataset, and the remaining unselected feature columns are removed. This operation unifies and reduces the feature dimensions of all datasets to 18 dimensions, constructing a cross-domain feature space with the same structure and corresponding physical meaning. This ensures consistency between subsequent model testing and transfer learning input data while eliminating redundant and noisy features.

[0039] By using the filtered features to fuse Bagging and Boosting algorithms, a soft-voting heterogeneous ensemble model containing four heterogeneous base classifiers is constructed, and the source domain fault diagnosis model is trained.

[0040] The specific structure and function of each base classifier are as follows: 1. Random Forest (RF): A Bagging-type algorithm, it sets the number of decision trees to 100, with no maximum depth limit. By introducing sample perturbation, it effectively reduces the variance of the model and improves the stability of acoustic emission signals from infill bodies.

[0041] 2. Extremely Random Trees (ET): This belongs to the Bagging class of algorithms, with a maximum of 100 decision trees. By maximizing the randomness of node splitting, it enhances the model's robustness against transient noise in acoustic emission signals.

[0042] 3. XGBoost (XGB): This is a boosting algorithm with a learning rate of 0.1, a maximum depth of 6, and a multi-class softmax objective function. By fitting the residuals, it effectively reduces model bias and improves classification accuracy for difficult samples.

[0043] 4. LightGBM (LGB): This is a Boosting algorithm with a learning rate of 0.1, 31 leaf nodes, and a GOSS strategy. This model aims to efficiently process large-scale acoustic emission signals and improve computational efficiency.

[0044] A soft voting strategy is employed to fuse the prediction results of the four base classifiers. Specifically, the prediction probability matrix of each base classifier for the current sample belonging to a certain destruction stage is calculated, and then fused according to a preset weight vector. We perform a weighted summation to obtain the final predicted probability: In the formula, For the final predicted probability, , , and They are respectively , , and Base classifier predicts samples Category The probability, , , and They are respectively , , and The weights of the base classifier. To balance the bias and variance of the source domain fault diagnosis model, the weights are set according to the principle that the weights of Boosting algorithms are higher than those of Bagging algorithms (i.e., ...). ),Pick The category with the highest probability is taken as the final fault diagnosis result of the source domain filling body. For example... Figure 4 The confusion matrix shown indicates that the source domain fault diagnosis model achieves an overall accuracy of 95.36% on the source domain test set, demonstrating significant effectiveness in identifying key damage stages.

[0045] S4. Acquire acoustic emission signals of the target domain filling body to obtain the target domain dataset, calculate the maximum mean difference between the target domain dataset and the source domain dataset. If the maximum mean difference exceeds the preset threshold or the test accuracy of the source domain fault diagnosis model on the target domain dataset is lower than the preset baseline value, it is determined that there is a significant domain shift and the transfer learning process is triggered.

[0046] This step aims to establish a quantitative set of criteria to determine whether the source domain fault diagnosis model developed in the laboratory can be directly applied to the new monitoring environment. The specific steps are as follows: Based on the methods described in steps S1 and S2, acoustic emission signals of the target domain filling body are collected to construct a target domain dataset. To clarify the physical background of the domain shift, in this embodiment, the source domain feature space is set to correspond to the first mix proportion condition (ash-sand ratio 1:4, slurry mass concentration 75%), and the target domain feature space is set to correspond to the second mix proportion condition (ash-sand ratio 1:8, slurry mass concentration 72%). Due to the decrease in the ash-sand ratio and the change in concentration, the pore structure and bonding strength inside the filling body change, resulting in a nonlinear drift in the statistical law of the acoustic emission signal.

[0047] To quantitatively assess this distributional difference, the source domain dataset was calculated. With the target domain dataset The maximum mean difference (MMD) between two probability distributions. MMD is a statistic that measures the distance between two probability distributions in the reproducing kernel Hilbert space (RKHS), and its formula is: In the formula, and These represent the number of samples in the source domain filling body and the target domain filling body, respectively. and The source domain filling bodies are respectively The first sample and the target domain filling body The feature vector of each sample (containing key physical features such as energy, count, and frequency); This represents a nonlinear mapping function that maps the original feature space (source domain feature space or target domain feature space) to the reproducing kernel Hilbert space (RKHS); The norm in the RKHS space is represented. To capture higher-order statistical information of acoustic emission characteristics, a Gaussian kernel function is used. The calculated MMD value between the source and target domain datasets in this embodiment is 0.0297. This value is significantly greater than 0 (the preset threshold), indicating that the characteristic distribution of the acoustic emission signal of the filling material has significantly shifted under the new environment, and the source and target domain datasets no longer satisfy the independent and identically distributed assumption.

[0048] While calculating the MMD, a zero-shot baseline test was performed to evaluate the model's generalization ability. Specifically, the source domain fault diagnosis model trained in step S3 based on the first matching ratio (1:4, 75%) was directly applied to the target domain dataset with the second matching ratio (1:8, 72%) for identification testing. The test results showed that the accuracy of the source domain fault diagnosis model was only 44.27% (e.g., ...). Figure 5 As shown in the figure, the values ​​were lower than the preset baseline, and the rate of missed diagnoses at each physical stage increased significantly.

[0049] In summary, the MMD value (0.0297) indicates a significant statistical difference in feature distribution, and the accuracy of the zero-sample baseline test (44.27%) demonstrates that the nonlinear shift caused by mix proportion fluctuations has a significant impact on traditional single models, far below the standards for engineering applications. Therefore, a severe domain shift is identified, and the subsequent transfer learning process is automatically triggered.

[0050] S5. Construct a hybrid dataset based on the target domain dataset and the source domain dataset. Using this hybrid dataset, update the source domain fault diagnosis model using a weighted fine-tuning transfer strategy to obtain the final infill body instability stage identification model. This model is then used to achieve intelligent identification of the infill body stages. Specifically, this includes: Constructing a hybrid training dataset: from the target domain dataset A small number of samples, ranging from 5% to 30%, are randomly selected and manually or with auxiliary annotation to serve as the training set for the target domain. The remaining data was then used as the target domain test set for validation. Subsequently, the full source domain dataset was... and Mixing, forming a mixed dataset .

[0051] To address the dominance effect caused by the source domain dataset being significantly larger than the target domain training set, and the negative transfer effect resulting from the bias in the source domain data distribution, this embodiment employs an instance-level weighting strategy. Specifically, when calculating the loss function during model training, different weights are assigned to samples from different sources: the target domain infill sample weights... :set up The target domain infill samples are assigned the highest weight, forcing the model to preferentially fit the characteristic distribution of "weak bonding and early-onset crack signals" in the target domain infill, ensuring the model's sensitivity to new operating conditions. Source domain infill sample weights. :set up By assigning low weights to the source domain filling samples, they exist only as regularization terms or background knowledge, preventing the model from overfitting on small samples while retaining the memory of the general laws of acoustic emission signals.

[0052] Using weighted mixed data The source domain fault diagnosis model in step S3 is updated through hierarchical fine-tuning. Considering the advantages of Boosting algorithms in fitting residuals and the stability of Bagging algorithms, a differentiated update strategy is adopted: Boosting base learner update: Fine-tuning parameters for XGBoost and LightGBM base learners. Explicitly introducing sample weights into the training objective function. Its weighted objective function It can be represented as: In the formula, For loss function, For regularization terms, For mixed datasets The Middle The true label of each sample For the model to the first The model predicts the values ​​for each sample. By minimizing the weighted objective function, the model can quickly correct for nonlinear biases caused by changes in the proportions.

[0053] Bagging base learner preservation: The parameters of RandomForest and ExtraTrees are kept unchanged or fine-tuned only with a very low learning rate to maintain the model's robust extraction ability of basic features of acoustic emission signals and prevent catastrophic forgetting during small-sample fine-tuning. Training parameter settings: The number of fine-tuning iterations is set to 50, and a learning rate decay strategy is adopted, so that the learning rate gradually decreases to 10% of the initial value as the number of iterations increases. This helps the model to finely search in the parameter space and eventually converge to the optimal solution that fits the target domain filling condition.

[0054] The final source domain fault diagnosis model performance is output, resulting in the final instability stage identification model for the filling body. This model is then used to achieve intelligent identification of the filling body stages. Experimental results show that this weighted fine-tuning strategy achieves high-precision cross-domain identification with low annotation costs: by extracting only 10% of the small sample data from the 1:8 mix ratio (target domain) database and performing weighted fine-tuning on the original 1:4 mix ratio (source domain) model, the model can successfully capture the weak cementation damage characteristics unique to the 1:8 mix ratio. Figure 5 As shown, when only 5% of the target domain samples were used for fine-tuning, the model's recognition accuracy jumped from 44.27% at the zero-sample baseline to 73.40%; when the target domain label ratio increased to 30%, the model's recognition accuracy for each stage of the second-ratio filling body further exceeded 90.19%. This indicates that the method successfully achieved high-precision cross-domain recognition of the instability stage of the filling body with extremely low data annotation costs.

[0055] The above results demonstrate that the method proposed in this invention, based on multi-model ensemble and transfer learning, can effectively solve the problem of inconsistent data distribution. It can achieve high-precision identification of the destruction stage of filling bodies with only a very small number of labeled samples, and has extremely high accuracy, robustness and engineering applicability.

[0056] The method described in this invention is applicable to stability monitoring and failure stage identification of cemented backfill bodies, paste backfill bodies, tailings cemented backfill bodies, and rock-backfill body composite structures in deep underground engineering.

[0057] In summary, this invention discloses an intelligent identification method for the filling stage based on multi-model ensemble and transfer learning. It automates the processing of source domain data through an adaptive annotation algorithm and achieves efficient model adaptation from a standard laboratory environment to the target domain through the organic integration of ensemble learning and transfer learning techniques. This method can be automated through computer programs, significantly reducing reliance on large amounts of manually labeled data from the target monitoring environment and exhibiting high operational efficiency. This invention not only has significant social value, effectively improving the early warning level of mine dynamic disasters, but also brings significant economic benefits, reducing monitoring costs. Furthermore, it technically overcomes the bottleneck of traditional acoustic emission monitoring models in solving the domain offset problem.

[0058] Example 2 Based on the same inventive concept, the present invention also provides an intelligent recognition system for filling body stages based on multi-model integration and transfer learning, used to implement the methods described in the foregoing embodiments. The system includes: a data acquisition and extraction module, a stage labeling module, a screening and construction module, an offset detection module, and a model update module. The acquisition and extraction module is used to acquire the acoustic emission signal of the source domain filling body using the constructed filling body acoustic emission monitoring system and perform noise reduction processing. Based on the noise-reduced acoustic emission signal, multi-dimensional feature vectors are extracted to construct the original feature dataset. The stage labeling module is used to build an adaptive labeling model based on the differential evolution algorithm, automatically identify the three critical time breakpoints of the four stages of the source domain filling failure process, generate stage labels for the source domain filling failure process, and obtain the source domain dataset. The filtering module is used to filter features in the original feature dataset based on the source domain dataset using a leak-proof feature optimization mechanism. The filtered features are then used to construct a soft-voting heterogeneous ensemble model by fusing Bagging and Boosting algorithms, and the source domain fault diagnosis model is trained. The offset detection module is used to collect the acoustic emission signal of the target domain filling body to obtain the target domain dataset, calculate the maximum mean difference between the target domain dataset and the source domain dataset, and if the maximum mean difference exceeds the preset threshold or the test accuracy of the source domain fault diagnosis model on the target domain dataset is lower than the preset baseline value, it is determined that there is a significant domain offset and the transfer learning process is triggered. The model update module is used to construct a hybrid dataset based on the target domain dataset and the source domain dataset. Based on the hybrid dataset, a weighted fine-tuning migration strategy is used to update the source domain fault diagnosis model to obtain the final infill body instability stage identification model. The infill body instability stage identification model is used to realize intelligent identification of the infill body stage.

[0059] Furthermore, in this embodiment, the multidimensional feature vector includes: Basic parameter characteristics: ring count, energy, amplitude, duration, rise time; Frequency domain characteristics: center frequency, peak frequency, average frequency, and frequency centroid; Waveform characteristics: RA value, AF value; Evolutionary characteristics: the cumulative value of basic parameters over time, the rate of change, and the event density.

[0060] Furthermore, in this embodiment, the method for constructing an adaptive annotation model based on the differential evolution algorithm includes: ; In the formula, Let be the objective function. There are three critical time breakpoints. For the first k Sum of squared fitting residuals within each stage The penalty coefficient is... P This is a physical constraint penalty term.

[0061] Furthermore, in this embodiment, the method for calculating the maximum mean difference between the target domain dataset and the source domain dataset includes: ; In the formula, For the maximum mean difference, and These represent the source domain dataset and the target domain dataset, respectively. and These represent the number of samples in the source and target domains, respectively. and The source domain is respectively The nth sample and the target domain The feature vector of each sample Let represent the mapping function to the reproducing kernel Hilbert space. denoted as the norm in the regenerated kernel Hilbert space.

[0062] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for intelligent identification of infill body stages based on multi-model ensemble and transfer learning, characterized in that, The method includes: The constructed acoustic emission monitoring system for infill bodies was used to collect acoustic emission signals from the source domain infill bodies and perform noise reduction processing. Based on the noise-reduced acoustic emission signals, multi-dimensional feature vectors were extracted to construct the original feature dataset. An adaptive labeling model based on differential evolution algorithm is constructed to automatically identify three critical time breakpoints in the four stages of the source domain filling failure process, generate stage labels for the source domain filling failure process, and obtain the source domain dataset. Based on the source domain dataset, a leakage prevention feature optimization mechanism is used to filter the features in the original feature dataset. The filtered features are then used to construct a soft voting heterogeneous ensemble model by fusing Bagging and Boosting algorithms, and the source domain fault diagnosis model is trained. Acoustic emission signals of the target domain filling body are collected to obtain the target domain dataset. The maximum mean difference between the target domain dataset and the source domain dataset is calculated. If the maximum mean difference exceeds a preset threshold or the test accuracy of the source domain fault diagnosis model on the target domain dataset is lower than the preset baseline value, a significant domain shift is determined and the transfer learning process is triggered. A hybrid dataset is constructed based on the target domain dataset and the source domain dataset. Based on the hybrid dataset, a weighted fine-tuning migration strategy is used to update the source domain fault diagnosis model to obtain the final infill body instability stage identification model. The infill body instability stage identification model is used to realize intelligent identification of the infill body stage.

2. The method according to claim 1, characterized in that, Multidimensional feature vectors include: basic parameter features, frequency domain features, waveform features, and evolutionary features; Basic parameter characteristics: ring count, energy, amplitude, duration, and rise time; Frequency domain characteristics: center frequency, peak frequency, average frequency, and frequency centroid; Waveform characteristics: RA value and AF value; Evolutionary characteristics: the cumulative value of basic parameters over time, the rate of change, and the event density.

3. The method according to claim 1, characterized in that, Methods for constructing adaptive annotation models based on differential evolution algorithms include: ； In the formula, Let be the objective function. There are three critical time breakpoints. For the first k Sum of squared residuals within each stage The penalty coefficient is... P This is a physical constraint penalty term.

4. The method according to claim 1, characterized in that, Methods for calculating the maximum mean difference between the target domain dataset and the source domain dataset include: ； In the formula, For the maximum mean difference, and These represent the source domain dataset and the target domain dataset, respectively. and These represent the number of samples in the source and target domains, respectively. and The source domain is respectively The nth sample and the target domain The feature vector of each sample Let represent the mapping function to the reproducing kernel Hilbert space. denoted as the norm in the regenerated kernel Hilbert space.

5. An intelligent recognition system for infill body stages based on multi-model ensemble and transfer learning, the system being used to implement the method described in any one of claims 1-4, characterized in that, The system includes: an acquisition and extraction module, a stage annotation module, a filtering and construction module, an offset detection module, and a model update module; The acquisition and extraction module is used to acquire the acoustic emission signal of the source domain filling body using the constructed filling body acoustic emission monitoring system and perform noise reduction processing. Based on the noise-reduced acoustic emission signal, multi-dimensional feature vectors are extracted to construct the original feature dataset. The stage labeling module is used to build an adaptive labeling model based on the differential evolution algorithm, automatically identify the three critical time breakpoints of the four stages of the source domain filling failure process, generate stage labels for the source domain filling failure process, and obtain the source domain dataset. The filtering module is used to filter features in the original feature dataset based on the source domain dataset using a leak-proof feature optimization mechanism. The filtered features are then used to construct a soft-voting heterogeneous ensemble model by fusing Bagging and Boosting algorithms, and the source domain fault diagnosis model is trained. The offset detection module is used to collect the acoustic emission signal of the target domain filling body to obtain the target domain dataset, calculate the maximum mean difference between the target domain dataset and the source domain dataset, and if the maximum mean difference exceeds the preset threshold or the test accuracy of the source domain fault diagnosis model on the target domain dataset is lower than the preset baseline value, it is determined that there is a significant domain offset and the transfer learning process is triggered. The model update module is used to construct a hybrid dataset based on the target domain dataset and the source domain dataset. Based on the hybrid dataset, a weighted fine-tuning migration strategy is used to update the source domain fault diagnosis model to obtain the final infill body instability stage identification model. The infill body instability stage identification model is used to realize intelligent identification of the infill body stage.

6. The system according to claim 5, characterized in that, Multidimensional feature vectors include: basic parameter features, frequency domain features, waveform features, and evolutionary features; Basic parameter characteristics: ring count, energy, amplitude, duration, and rise time; Frequency domain characteristics: center frequency, peak frequency, average frequency, and frequency centroid; Waveform characteristics: RA value and AF value; Evolutionary characteristics: the cumulative value of basic parameters over time, the rate of change, and the event density.

7. The system according to claim 5, characterized in that, Methods for constructing adaptive annotation models based on differential evolution algorithms include: ； In the formula, Let be the objective function. There are three critical time breakpoints. For the first k Sum of squared residuals within each stage The penalty coefficient is... P This is a physical constraint penalty term.

8. The system according to claim 5, characterized in that, Methods for calculating the maximum mean difference between the target domain dataset and the source domain dataset include: ； In the formula, For the maximum mean difference, and These represent the source domain dataset and the target domain dataset, respectively. and These represent the number of samples in the source and target domains, respectively. and The source domain is respectively The nth sample and the target domain The feature vector of each sample Let represent the mapping function to the reproducing kernel Hilbert space. denoted as the norm in the regenerated kernel Hilbert space.

Images