Filling body life prediction method and system based on energy constraint fusion deep learning

Abstract

The present application relates to a filling body life prediction method and system based on energy constraint fusion deep learning, belonging to the technical field of filling body structure health monitoring and life prediction, the method comprising: collecting the transient elastic wave signal of the filling body sample; performing multi-stage data cleaning processing on the transient elastic wave signal; based on the cleaned data, using a physical damage evolution model based on cumulative energy normalization to calculate damage stage characteristics; based on the cleaned data, constructing multi-scale features; inputting the original waveform sequence features, multi-scale features and damage stage characteristics in the sliding window into a Transformer deep neural network model based on multi-source information fusion for model training to obtain a prediction model; and using the prediction model to predict the remaining life of the filling body. The present application can accurately predict the remaining service life of the filling body structure in real time, and has important theoretical significance and practical application value.

Description

Technical Field

[0001] This invention relates to the field of infill structure health monitoring and life prediction technology, and in particular to a method and system for infill life prediction based on energy-constrained fusion deep learning. Background Technology

[0002] As the core support structure in backfill mining, the backfill body is composed of a mixture of tailings, waste rock, and cementing materials. Its mechanical properties and stability directly affect the safety of the goaf and the control of surface subsidence. Due to the large number of micropores and highly heterogeneous component distribution within the backfill body, it is highly susceptible to nonlinear evolution processes involving the initiation, propagation, and eventual macroscopic instability and failure under the high ground stress and complex environment of deep mining. Accurately predicting the remaining useful life (RUL) of the backfill body is crucial for developing reasonable maintenance strategies and preventing disasters.

[0003] Acoustic emission (AE), as a dynamic nondestructive testing technique, can capture transient elastic waves released from internal damage and fracture of filling materials in real time, and is widely used to monitor their stability. Current methods for predicting remaining lifetime mainly suffer from the following problems: Traditional empirical formula-based methods rely on manual judgment and are highly subjective. Researchers often employ parametric analysis methods, using parameters such as ring count, energy, and amplitude to calculate b-values, Ib-values, or RA-AF correlation values ​​to assess damage status and infer remaining life. However, such methods require preset fixed empirical thresholds. Due to the diverse composition of mine backfill materials and complex stress conditions, fixed thresholds are difficult to adapt to the damage evolution patterns under different cementation strengths or loading rates.

[0004] Existing deep learning methods lack physical constraints and suffer from poor model interpretability. Although deep learning models such as convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) have been introduced into the field of life prediction in recent years, these methods often treat life prediction as a purely data fitting problem, ignoring the physical mechanisms of material damage evolution, resulting in insufficient reliability and interpretability of the models in engineering practice.

[0005] Insufficient utilization of feature information. Existing methods typically use only a single type of feature for prediction, failing to effectively integrate multi-source heterogeneous data such as waveform sequence features, statistical features, and damage stage information, resulting in low information utilization and limited prediction accuracy.

[0006] The damage evolution process exhibits nonlinear characteristics, and traditional linear assumptions lead to large prediction errors. The damage accumulation process of the filling material has obvious nonlinear characteristics, with slow initial damage growth followed by a sharp acceleration near failure. Traditional linear damage assumptions cannot accurately describe this evolutionary pattern.

[0007] Therefore, there is an urgent need for a method to predict the remaining lifetime of infill bodies that can integrate physical constraints and deep learning. Summary of the Invention

[0008] The purpose of this invention is to propose a method and system for predicting the lifespan of infill bodies based on energy-constrained deep learning. The aim is to establish a physical damage model based on accumulated energy and combine it with deep learning technology to analyze the temporal characteristics of acoustic emission signals of infill bodies, thereby achieving high-precision intelligent prediction of the remaining lifespan of infill bodies and providing a scientific basis for structural health monitoring, reinforcement and maintenance.

[0009] To achieve the above objectives, the present invention provides the following solution: A method for predicting the lifetime of infill bodies based on energy-constrained fusion deep learning includes: Acquire transient elastic wave signals from the filling sample; The transient elastic wave signal is subjected to multi-level data cleaning processing; Based on the cleaned data, the damage stage characteristics are calculated using a physical damage evolution model based on cumulative energy normalization; wherein, the damage stage characteristics include: physical damage degree and remaining lifetime; Based on the cleaned data, multi-scale features are constructed; wherein, the multi-scale features include: RA value features, AF value features, and multi-window statistics; The original waveform sequence features, the multi-scale features, and the damage stage features within the sliding window are input into a Transformer deep neural network model based on multi-source information fusion to train the model and obtain a prediction model. The prediction model is used to predict the remaining life of the filling material.

[0010] Optionally, a unified parsing rule can be established to address the non-standard numerical representations mixed in the original data; Remove blank data columns with zero information content, mark infinite values ​​caused by numerical overflow as invalid, and identify statistical outliers based on the interquartile range criterion and perform truncation. The median of each parameter column is used as the filler value for missing values; After processing, perform data validity verification to check the sequence length and the dynamic range of key acoustic emission parameters.

[0011] Optionally, calculating the damage stage characteristics includes: Calculate the cumulative energy sequence : ; in, For the first The instantaneous energy value of each sampling point, where t is the time of the sampling point; Calculate background energy : ; in, , The total length of the data. Indicates rounding down; Determine peak energy : ; in, This is the moment when the stress reaches its peak. Calculate real-time physical damage : ; in, To prevent tiny quantities with a denominator of zero; Calculate the percentage of remaining lifespan : .

[0012] Optionally, constructing the multi-scale features includes: Calculate the RA value characteristics: ; in, Indicates the rise time of the acoustic emission signal. Indicates the amplitude of the acoustic emission signal. To prevent tiny quantities with a denominator of zero; Calculate the AF value characteristics: ; in, The ring count represents the acoustic emission signal. Indicates the duration of the acoustic emission signal; Calculate the multi-window statistics: Based on the multi-scale sliding window technique, perform rolling statistical analysis on the energy, amplitude, RA value characteristics and AF value characteristics respectively, and calculate the mean, standard deviation and first difference slope in each window; A high-dimensional feature pool is constructed based on the RA value features, AF value features and multi-window statistics. The SelectKBest algorithm is used in combination with the F value of variance analysis to evaluate the correlation between each feature and the remaining lifetime, and multiple features with preset contribution are selected. RobustScaler was used to perform robust standardization on the filtered features; The feature sequence after robust standardization is divided into three-dimensional tensor samples using the sliding window method, and the label of each sample is the remaining lifetime value at the end of the window.

[0013] Optionally, the original waveform sequence features within the sliding window, the multi-scale features, and the damage stage features are input into a Transformer deep neural network model based on multi-source information fusion, including: The original waveform sequence features, the multi-scale features, and the damage stage features within the sliding window are aggregated along the feature dimension through a splicing operation, and then a unified fused feature representation is generated through a fusion network; wherein, the fusion network includes: layer normalization and linear transformation; The fused feature representation is input into the Transformer encoder for temporal modeling. The Transformer encoder uses a standard multi-head self-attention mechanism, and a learnable sinusoidal position code is added before inputting into the encoder to preserve temporal position information. The encoder output is then subjected to temporal aggregation through an attention pooling layer. The pooled feature vectors are input into the prediction head for final remaining lifetime prediction.

[0014] Optionally, the attention pooling layer performs a weighted summation of features at each time step using a learnable attention weight vector.

[0015] Optionally, the prediction head adopts a multilayer perceptron structure, containing two fully connected layers, with GELU activation function and Dropout regularization used in between, and the final output dimension is 1, corresponding to the predicted percentage of remaining lifetime.

[0016] Optionally, model training includes: Logarithmic transformation of physical damage level is applied to optimize the training process; After the model output is obtained, it is restored to the physical damage level through inverse transformation, and then the remaining lifetime is calculated. The model training uses the MSE loss function; The MSE loss function is: ; in, Labels are for the actual remaining lifespan. These are the model's predicted values. The MSE loss function is... This represents the total length of the data.

[0017] An energy-constrained deep learning-based infill lifetime prediction system includes: The data acquisition module is used to acquire transient elastic wave signals of the filling sample; The preprocessing module is used to perform multi-level data cleaning processing on the transient elastic wave signal; The calculation module, based on the cleaned data, uses a physical damage evolution model based on cumulative energy normalization to calculate the damage stage characteristics; wherein, the damage stage characteristics include: physical damage degree and remaining lifetime; The feature construction module is used to construct multi-scale features based on the cleaned data; wherein, the multi-scale features include: RA value features, AF value features, and multi-window statistics; The model training module is used to use the remaining lifetime as a label, input the original waveform sequence features, the multi-scale features and the damage stage features within the sliding window, the Transformer deep neural network model based on multi-source information fusion, and perform model training to obtain a prediction model. The prediction output module is used to predict the remaining life of the filling material using the prediction model.

[0018] Optionally, the data acquisition module includes: The signal sensing unit is used to capture high-frequency elastic waves generated by micro-fractures inside the filling body using a broadband piezoelectric ceramic sensor. The sensor is arranged in a symmetrical configuration, with one sensor installed on each side of the sample. An acoustic coupling agent is coated between the sensor and the sample surface. A signal trigger threshold is set, and data recording is only initiated when the signal amplitude exceeds the trigger threshold. Stress-strain data is collected synchronously, and the correlation between acoustic emission signals and mechanical responses is realized through timestamp alignment technology. The signal conditioning unit is used to improve the signal-to-noise ratio of weak signals using a low-noise preamplifier; A digital acquisition unit is used to convert analog signals into digital signals.

[0019] The beneficial effects of this invention are as follows: 1. Strong physical interpretability: The proposed physical damage model based on cumulative energy normalization uses the average of the top 5% of the cumulative energy sequence as the background energy and the cumulative energy corresponding to the peak stress point as the peak energy, which has clear physical meaning. The damage degree D(t) directly reflects the evolution process of energy accumulation and release inside the material, and the remaining lifetime RUL(t) is complementary to the damage degree, which is in line with the basic laws of materials mechanics.

[0020] 2. Deep fusion of multi-source information: The model simultaneously integrates original waveform features, manual statistical features (RA value, AF value, multi-window statistics) and damage stage features, making full use of multi-scale and multi-level information to improve prediction accuracy.

[0021] 3. Scientific training optimization strategy: Logarithmic transformation is used to solve the problem of uneven distribution of damage values, so that the model can simultaneously focus on the evolution of the initial damage stage and the critical failure stage.

[0022] 4. High prediction accuracy: Experimental results show that the method achieves a lifetime prediction R² of 0.9851, MAE of 1.47%, and RMSE of 2.19% on the validation set, which is significantly better than the traditional method.

[0023] 5. High adaptability: Based on data-driven feature selection and model training strategies, it can automatically adapt to the damage evolution of filling materials under different ratios and loading conditions.

[0024] 6. Automated implementation: This method can be automated through computer programs, which has high operational efficiency and repeatability.

[0025] In summary, this invention provides an innovative method and system for predicting the lifetime of infill structures based on energy-constrained fusion deep learning. This method can predict the remaining lifetime of infill structures in real time and accurately, and has significant theoretical and practical application value. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be 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.

[0027] Figure 1 This is a schematic diagram of the infill lifetime prediction method based on energy-constrained fusion deep learning according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the physical damage model based on accumulated energy according to an embodiment of the present invention, illustrating... , , and Relationship; Figure 3 The flowchart for constructing multi-scale feature engineering in an embodiment of the present invention illustrates the extraction and filtering process of RA value, AF value, and multi-window statistical features. Figure 4 This is a diagram of the multi-source information fusion Transformer model architecture according to an embodiment of the present invention, which shows the internal structure of the network in detail; Figure 5 The graph shown here is a performance evaluation chart of the model prediction in an embodiment of the present invention, which illustrates the regression analysis results and error distribution of the actual remaining lifetime and the predicted remaining lifetime. Figure 6This is a comparison chart of the remaining lifetime prediction effects of an embodiment of the present invention, showing the degree of fit between the actual value and the predicted value. Detailed Implementation

[0028] 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.

[0029] 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.

[0030] like Figure 1 As shown, this embodiment proposes a method for predicting the lifetime of infill bodies based on energy-constrained fusion deep learning, including: S1. Acquire the transient elastic wave signal of the filling sample; S2. Perform multi-level data cleaning processing on the transient elastic wave signal; S3. Based on the cleaned data, the damage stage characteristics are calculated using a physical damage evolution model based on cumulative energy normalization; wherein, the damage stage characteristics include: physical damage degree and remaining lifetime; S4. Based on the cleaned data, construct multi-scale features; wherein, the multi-scale features include: RA value features, AF value features, and multi-window statistics; S5. Input the original waveform sequence features, the multi-scale features, and the damage stage features within the sliding window into the Transformer deep neural network model based on multi-source information fusion, train the model, and obtain the prediction model. S6. Using the prediction model, predict the remaining life of the filling body.

[0031] This embodiment establishes a multi-channel acoustic emission real-time monitoring platform to continuously acquire transient elastic wave signals generated by the filling material sample under uniaxial compression loading conditions and performs a multi-level data cleaning process. A physical damage evolution model based on cumulative energy normalization is established, using the average of the top 5% of the cumulative energy sequence as the background energy and the cumulative energy corresponding to the peak stress point as the peak energy to calculate the real-time damage degree and remaining life percentage. A multi-scale feature engineering system is constructed to extract RA values, AF values, and multi-window sliding statistical features. A multi-source information fusion Transformer deep neural network model integrating waveform features, manual features, and stage features is constructed to achieve high-precision prediction of the remaining life of the filling material.

[0032] In this embodiment, S1 builds a multi-channel acoustic emission real-time monitoring platform to continuously collect and store the transient elastic wave signals generated by the filling sample under uniaxial compression loading conditions.

[0033] S2. Perform a multi-level data cleaning process on the original acoustic emission parameter sequence, including format normalization, outlier detection and correction, and intelligent imputation of missing data. The data preprocessing process performs the following operations in sequence: To address the non-standard numerical representations mixed in the original data, a unified parsing rule is established: for numerical values ​​containing comparison operators (such as "<", ">", "≤", "≥"), the numerical part is extracted; for interval expressions (such as "10-20"), the midpoint of the interval is calculated as the representative value; for percentage forms, they are automatically converted to the corresponding decimals; for extremely small or extremely large numerical values ​​expressed in exponential notation, they are parsed into standard floating-point number format.

[0034] Remove blank data columns with zero information content; mark infinite values ​​caused by numerical overflow as invalid; identify statistical outliers in each parameter column based on the interquartile range (IQR) criterion, and truncate extreme values ​​that exceed the reasonable range.

[0035] An imputation strategy based on local distribution characteristics is adopted, using the median of each parameter column as the missing value to fill the missing value. Compared with mean imputation, this method has stronger anti-interference ability and can effectively avoid the influence of extreme values ​​on the imputation result.

[0036] After preprocessing, perform data validity verification: check whether the sequence length meets the minimum requirements for subsequent analysis, verify whether the key acoustic emission parameters (energy, amplitude, etc.) have sufficient dynamic range, and ensure the integrity and reliability of the input data.

[0037] S3. Establish a physical damage evolution model based on cumulative energy normalization, and calculate the real-time damage degree D(t) and remaining lifetime RUL(t). The principle of the physical damage model is as follows: Figure 2 As shown. This model is based on the physical correlation between acoustic emission energy accumulation and material damage evolution, and the specific steps are as follows: First, calculate the cumulative energy sequence: ; in, For the first Instantaneous energy value at each sampling point (unit: mV·mS).

[0038] Secondly, calculate the background energy. The average value of the first 5% of data points on the cumulative energy curve is taken as the background energy, representing the background energy level of the material in its initial state. ; in, , The total length of the data. This indicates rounding down to the nearest integer.

[0039] Next, determine the peak energy. Acquire stress data during the loading process and record the moment when the stress reaches its peak. Take the cumulative energy value corresponding to that moment: ; If stress data is unavailable, the alternative approach is to use the cumulative energy value corresponding to the point when the normalized cumulative energy first reaches 85%. .

[0040] Then, calculate the real-time physical damage degree: ; in, To prevent tiny quantities with a denominator of zero, the value is taken as... .

[0041] Physical constraints are applied to the damage level to ensure that its value is within the range of [0, 1]. This constraint ensures that the damage degree is close to 0 at the initial moment and close to 1 at the peak stress, which is in line with the physical laws of material damage evolution.

[0042] Finally, calculate the remaining lifespan percentage: ; The remaining lifetime and the degree of damage are complementary. When the degree of damage is 0, the remaining lifetime is 100%; when the degree of damage is 1, the remaining lifetime is 0%, indicating that the material is about to undergo macroscopic failure.

[0043] The remaining lifetime is used as a label. The model uses the damage level calculated by the physical model (such as the initial stage, stable stage, and accelerated failure stage) as input information to the Transformer, rather than as the prediction target. The physical damage level is input into the neural network as a physical constraint to help the model understand the current physical state.

[0044] S4. Construct a multi-scale feature engineering system to extract the time-domain features and engineering parameter features of the acoustic emission signal. The feature engineering construction process is as follows: Figure 3 As shown. First, physical correlation characteristics are constructed based on the acoustic emission mechanism: Constructing RA value characteristics to characterize crack type: ; In the formula, The rise time of the acoustic emission signal is expressed in μs. This indicates the amplitude of the acoustic emission signal (unit: dB). To prevent minute values ​​with a denominator of zero, a larger RA value indicates that shear cracks are dominant, while a smaller RA value indicates that tensile cracks are dominant.

[0045] Constructing AF value features to characterize the signal frequency characteristics: ; In the formula, The ring count represents the acoustic emission signal. The AF value represents the duration of the acoustic emission signal (in μs). The AF value reflects the average frequency characteristics of the signal.

[0046] Subsequently, a multi-scale sliding window technique was introduced, with window sizes of 10 and 30, to perform rolling statistical analysis on energy, amplitude, RA value and AF value, and to calculate the mean, standard deviation and first difference slope within each window.

[0047] The formula for calculating the window mean is: ; The formula for calculating the window standard deviation is: ; The formula for calculating the slope of the first difference is (only for a window size of 10): ; in, For window size, Let be the feature value of the i-th sampling point.

[0048] Based on this, a high-dimensional feature pool containing original features and statistical features is constructed. The SelectKBest algorithm combined with the F-value (f_classif) of the analysis of variance is used to evaluate the correlation between each feature and the remaining lifetime label, and the 60 features with the highest contribution are automatically selected.

[0049] RobustScaler is used to perform robust normalization on the filtered features: ; in, , , These represent the first quartile, median, and third quartile of the feature, respectively. This method is robust to outliers and can effectively suppress the interference of outliers on the standardization results.

[0050] Finally, the time series, i.e., the feature sequence after robust normalization, is divided into three-dimensional tensor samples using the sliding window method, with a window length of 64 and a sliding step size of 4. The label of each sample is the remaining lifetime value at the end of the window, ensuring the consistency between the predicted target and the current state.

[0051] S5. Construct a Transformer deep neural network model based on multi-source information fusion.

[0052] Model architecture such as Figure 4 As shown, this model employs a multi-source information fusion strategy, simultaneously receiving three types of heterogeneous input features and performing deep fusion.

[0053] The first category consists of original waveform sequence features within a sliding window, containing multidimensional time-series data with a window length of 64. The original feature dimensions are mapped to a latent space representation of d=128 dimensions through a linear projection layer. The original waveform sequence features within the sliding window are obtained by segmenting the cleaned data using a sliding window technique with a window length of 64 and a sliding step size of 4, resulting in fixed-length three-dimensional tensor samples.

[0054] The second category consists of manually extracted statistical features, including RA values, AF values, and multi-window statistics, totaling 60 dimensions. These features are mapped to a d=128-dimensional space through a feature projection layer (containing two fully connected network layers and the GELU activation function).

[0055] The third category is the damage stage characteristics at the current moment, which adopts stage encoding based on physical damage degree D(t) and is mapped to a d=128-dimensional space through a stage projection layer. The damage stage is divided into three intervals according to the damage degree: initial stage (D<0.3), stable development stage (0.3≤D<0.7), and accelerated destruction stage (D≥0.7).

[0056] The three features are aggregated along the feature dimension through a concatenation operation, and then a unified fused feature representation is generated through a fusion network (which includes layer normalization and linear transformation).

[0057] The fused features are input into the Transformer encoder for temporal modeling. The encoder employs a standard multi-head self-attention mechanism, with 8 attention heads, 128 hidden layer dimensions, 256 feedforward network dimensions, and 4 stacked layers. A learnable sinusoidal positional code is added before the input to the encoder to preserve temporal positional information. The attention calculation formula is: ; Where Q, K, and V are the query, key, and value matrices, respectively. Let be the dimension of the key vector. The multi-head attention mechanism projects the input into multiple subspaces for parallel attention computation, then concatenates the outputs of each head and performs a linear transformation to obtain the final result.

[0058] The encoder output undergoes temporal aggregation via an attention pooling layer. Unlike traditional average pooling or max pooling, attention pooling uses a learnable attention weight vector to perform a weighted summation of features at each time step. ; ; in, Let w be the hidden state at time step t, and w be the learnable attention weight vector. These are the normalized attention coefficients. This mechanism enables the model to adaptively focus on the time steps most important for the prediction task.

[0059] The pooled feature vectors are input into the prediction head for final remaining lifetime prediction. The prediction head uses a multilayer perceptron structure, containing two fully connected layers, with GELU activation and Dropout regularization (dropout rate of 0.2) in between. The final output dimension is 1, corresponding to the predicted remaining lifetime percentage.

[0060] S6. The model is trained and optimized using a physical label dataset to achieve end-to-end intelligent prediction of the remaining life of the infill body. The performance curve during the training process is shown below. Figure 5 As shown. First, a logarithmic transformation is applied to the damage labels to optimize the training process. Because the damage evolution process is non-linear, with slow initial damage growth followed by a rapid acceleration near destruction, the damage values ​​are extremely unevenly distributed within the [0, 1] interval, with a large number of samples concentrated in the low-damage region. To address this issue, a logarithmic transformation is employed: ; This transformation maps the damage level from the [0, 1] interval to the [0, 1] interval, but makes the distribution more uniform.

[0061] After the model output is completed, it is restored to the physical damage level through inverse transformation: ; Then calculate the remaining lifetime: ; The model training uses the MSE loss function: ; in, Labels are for the actual remaining lifespan. These are the model's predicted values.

[0062] The optimizer uses AdamW, with specific parameter settings: the initial learning rate is... The weight decay coefficient is The beta coefficient is (0.9, 0.999); the batch size is 64. The learning rate scheduling uses the OneCycleLR strategy, with a maximum learning rate of... It can dynamically adjust the learning rate during training, first increasing and then decreasing it, which helps the model escape local optima. Gradient pruning is enabled to limit the gradient norm to within 1.0, preventing gradient explosion. An early stopping mechanism is introduced, using the validation set R² as the monitoring metric. When R² does not increase within 10 consecutive epochs, training automatically stops and the optimal model weights are saved, achieving an optimal balance between model performance and training efficiency.

[0063] This embodiment also proposes an infill lifetime prediction system based on energy-constrained fusion deep learning, including: The data acquisition module is used to acquire transient elastic wave signals of the filling sample; The preprocessing module is used to perform multi-level data cleaning processing on the transient elastic wave signal; The calculation module, based on the cleaned data, uses a physical damage evolution model based on cumulative energy normalization to calculate the damage stage characteristics; wherein, the damage stage characteristics include: physical damage degree and remaining lifetime; The feature construction module is used to construct multi-scale features based on the cleaned data; wherein, the multi-scale features include: RA value features, AF value features, and multi-window statistics; The model training module is used to use the remaining lifetime as a label, input the original waveform sequence features, the multi-scale features and the damage stage features within the sliding window, the Transformer deep neural network model based on multi-source information fusion, and perform model training to obtain a prediction model. The prediction output module is used to predict the remaining life of the filling material using the prediction model.

[0064] Specifically, in this embodiment, the data acquisition module consists of four parts: a signal sensing unit, a signal conditioning unit, a digitization acquisition unit, and a data processing terminal. The signal sensing unit uses a wideband piezoelectric ceramic sensor with an operating frequency covering the range of 50kHz to 400kHz, effectively capturing high-frequency elastic waves generated by micro-fractures within the filling material. The signal conditioning unit includes a low-noise preamplifier to improve the signal-to-noise ratio of weak signals, with a total gain set to 40dB. The digitization acquisition unit is responsible for converting analog signals into digital signals, with a sampling rate of no less than 10MHz to ensure complete recording of waveform details.

[0065] The sensors were arranged symmetrically, with one sensor installed on each opposite side of the specimen. An acoustic coupling agent was coated between the sensors and the specimen surface to ensure good acoustic contact. To eliminate the influence of environmental electromagnetic interference and mechanical noise, a signal trigger threshold was set; data recording was only initiated when the signal amplitude exceeded this threshold. Stress-strain data were acquired synchronously, and a timestamp alignment technique was used to achieve precise correlation between the acoustic emission signal and the mechanical response. Before the experiment, a standardized simulated acoustic emission source was used to calibrate the monitoring system, verifying the consistency and sensitivity of the responses of each channel.

[0066] Experimental Results and Analysis The experiment used acoustic emission data from multiple sets of uniaxial compression tests of infill bodies for validation. The dataset was divided into training and validation sets in an 8:2 ratio. Evaluation metrics included mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²), calculated using the following formulas: ; ; ; in, For the true value, For predicted values, is the mean of the true values, and N is the sample size.

[0067] Model prediction performance evaluation, such as Figure 5 As shown in the figure, this figure illustrates the regression analysis results of the actual remaining lifetime and the predicted remaining lifetime for the test set samples. Figure 5 The horizontal axis represents the actual remaining lifetime, the vertical axis represents the predicted remaining lifetime, and the diagonal dashed line represents the ideal prediction line. From... Figure 5 As can be seen, the data points are highly concentrated near the ideal prediction line, and the darker colored areas completely overlap with the diagonal. Figure 5 The edge histograms at the top and right respectively illustrate the probability distribution density of the true and predicted values. The distribution patterns of the two are highly consistent, further validating the model's predictive stability and reliability throughout its entire lifecycle. Statistical indicators show that the model of this invention performs excellently in lifecycle prediction tasks, with a coefficient of determination (R²) as high as 0.9851, a root mean square error (RMSE) of only 2.19%, and a mean absolute error (MAE) of 1.47%.

[0068] Predictive effect comparison Figure 6 As shown. Figure 6 This visually demonstrates the comparison between the predicted remaining lifespan curve (blue) and the actual curve (black) throughout the entire lifespan. From... Figure 6As can be seen, the predicted curve closely follows the downward trend of the actual curve, and even in a magnified view, the predicted values ​​can keenly capture the fluctuations in the actual lifespan. Combined with... Figure 5 The regression analysis results show that the model not only accurately predicts the overall trend, but also achieves a very high level of accuracy in specific numerical values ​​(R²>0.98), indicating that the model has good generalization ability and robustness.

[0069] To verify the superiority of the method in this embodiment, a comparative experiment was conducted with traditional empirical formula methods, BP neural networks, LSTM networks, and CNN-LSTM hybrid models. Experimental results show that: traditional empirical formula methods, due to their reliance on manual thresholds, have poor generalization ability, with a MAE as high as 15.23% and an R² of only 0.6834; BP neural networks are limited by their feature extraction capabilities, with an MAE of 8.56% and an R² of 0.8245; while LSTM networks can handle time-series data, they suffer from the vanishing gradient problem in long sequences, with an MAE of 6.23% and an R² of 0.8756; although the CNN-LSTM hybrid model performs relatively well (MAE of 4.45% and R² of 0.9123), its accuracy still has room for improvement due to the lack of physical constraints. In contrast, the method and system proposed in this embodiment, based on the fusion of accumulated energy physical constraints and deep learning, significantly outperforms the above methods in the remaining lifetime prediction task. Compared to the second-best performing CNN-LSTM hybrid model, the method of this invention reduces the MAE by approximately 67% and improves the R² by 7.3 percentage points, demonstrating extremely high accuracy and reliability.

[0070] This embodiment discloses a method and system for predicting the lifetime of infill bodies based on energy-constrained deep learning. By organically integrating physical damage models with deep learning technology, it achieves high-precision automatic prediction of the remaining lifetime of infill bodies. This method can be automated through computer programs, resulting in high operational efficiency. This invention not only has significant social value but also brings substantial economic benefits, while technically achieving a deep understanding and accurate prediction of the damage evolution process of infill bodies.

[0071] Experiments show that the method achieves a lifetime prediction R² of 0.9851 and a MAE of 1.47% on the validation set. This method possesses advantages such as strong physical interpretability, high prediction accuracy, and strong adaptability, and can be widely applied to health monitoring and lifetime early warning of mine backfill bodies, tailings dam bodies, and deep underground engineering structures.

[0072] 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 predicting the lifetime of infill bodies based on energy-constrained fusion deep learning, characterized in that, include: Acquire transient elastic wave signals from the filling sample; The transient elastic wave signal is subjected to multi-level data cleaning processing; Based on the cleaned data, the damage stage characteristics are calculated using a physical damage evolution model based on cumulative energy normalization; wherein, the damage stage characteristics include: physical damage degree and remaining lifetime; Calculating the characteristics of the damage stage includes: Calculate the cumulative energy sequence : ； in, For the first The instantaneous energy value of each sampling point, where t is the time of the sampling point; Calculate background energy : ； in, , The total length of the data. Indicates rounding down; Determine peak energy : ； in, This is the moment when the stress reaches its peak. Calculate real-time physical damage : ； in, To prevent tiny quantities with a denominator of zero; Calculate the percentage of remaining lifespan : ； Based on the cleaned data, multi-scale features are constructed; wherein, the multi-scale features include: RA value features, AF value features, and multi-window statistics; The original waveform sequence features, the multi-scale features, and the damage stage features within the sliding window are input into a Transformer deep neural network model based on multi-source information fusion to train the model and obtain a prediction model. The prediction model is used to predict the remaining life of the filling material.

2. The infill lifetime prediction method based on energy-constrained fusion deep learning according to claim 1, characterized in that, To address the non-standard numerical representations mixed in the original data, a unified parsing rule is established; Remove blank data columns with zero information content, mark infinite values ​​caused by numerical overflow as invalid, and identify statistical outliers based on the interquartile range criterion and perform truncation. The median of each parameter column is used as the missing value to fill in the missing value; After processing, perform data validity verification to check the sequence length and the dynamic range of key acoustic emission parameters.

3. The infill lifetime prediction method based on energy-constrained fusion deep learning according to claim 1, characterized in that, Constructing the multi-scale features includes: Calculate the RA value characteristics: ； in, Indicates the rise time of the acoustic emission signal. Indicates the amplitude of the acoustic emission signal. To prevent tiny quantities with a denominator of zero; Calculate the AF value characteristics: ； in, The ring count represents the acoustic emission signal. Indicates the duration of the acoustic emission signal; Calculate the multi-window statistics: Based on the multi-scale sliding window technique, perform rolling statistical analysis on the energy, amplitude, RA value characteristics and AF value characteristics respectively, and calculate the mean, standard deviation and first difference slope in each window; A high-dimensional feature pool is constructed based on the RA value features, AF value features and multi-window statistics. The SelectKBest algorithm is used in combination with the F value of variance analysis to evaluate the correlation between each feature and the remaining lifetime, and multiple features with preset contribution are selected. RobustScaler was used to perform robust standardization on the filtered features; The feature sequence after robust standardization is divided into three-dimensional tensor samples using the sliding window method, and the label of each sample is the remaining lifetime value at the end of the window.

4. The infill lifetime prediction method based on energy-constrained fusion deep learning according to claim 1, characterized in that, The original waveform sequence features within the sliding window, the multi-scale features, and the damage stage features are input into a Transformer deep neural network model based on multi-source information fusion, including: The original waveform sequence features, the multi-scale features, and the damage stage features within the sliding window are aggregated along the feature dimension through a splicing operation, and then a unified fused feature representation is generated through a fusion network; wherein, the fusion network includes: layer normalization and linear transformation; The fused feature representation is input into the Transformer encoder for temporal modeling. The Transformer encoder uses a standard multi-head self-attention mechanism. Before inputting the encoder, a learnable sinusoidal position code is added to preserve temporal position information. The encoder output is then subjected to temporal aggregation through an attention pooling layer. The pooled feature vectors are input into the prediction head for final remaining lifetime prediction.

5. The infill lifetime prediction method based on energy-constrained fusion deep learning according to claim 4, characterized in that, The attention pooling layer performs a weighted summation of features at each time step using a learnable attention weight vector.

6. The infill lifetime prediction method based on energy-constrained fusion deep learning according to claim 4, characterized in that, The prediction head adopts a multilayer perceptron structure, which contains two fully connected layers, uses the GELU activation function and Dropout regularization in between, and finally outputs a dimension of 1, corresponding to the predicted percentage of remaining lifetime.

7. The infill lifetime prediction method based on energy-constrained fusion deep learning according to claim 1, characterized in that, Model training includes: Logarithmic transformation of physical damage level is applied to optimize the training process; After the model output is obtained, it is restored to the physical damage level through inverse transformation, and then the remaining lifetime is calculated. The model training uses the MSE loss function; The MSE loss function is: ； in, Labels are for the actual remaining lifespan. These are the model's predicted values. The MSE loss function is... This represents the total length of the data.

8. A system for predicting the lifetime of infill bodies based on energy-constrained fusion deep learning, characterized in that, The system for implementing the method as described in any one of claims 1-7 comprises: The data acquisition module is used to acquire transient elastic wave signals of the filling sample; The preprocessing module is used to perform multi-level data cleaning processing on the transient elastic wave signal; The calculation module, based on the cleaned data, uses a physical damage evolution model based on cumulative energy normalization to calculate the damage stage characteristics; wherein, the damage stage characteristics include: physical damage degree and remaining lifetime; The feature construction module is used to construct multi-scale features based on the cleaned data; wherein, the multi-scale features include: RA value features, AF value features, and multi-window statistics; The model training module is used to use the remaining lifetime as a label, input the original waveform sequence features, the multi-scale features and the damage stage features within the sliding window, the Transformer deep neural network model based on multi-source information fusion, and perform model training to obtain a prediction model. The prediction output module is used to predict the remaining life of the filling material using the prediction model.

9. The infill lifetime prediction system based on energy-constrained fusion deep learning according to claim 8, characterized in that, The data acquisition module includes: The signal sensing unit is used to capture high-frequency elastic waves generated by micro-fractures inside the filling body using a broadband piezoelectric ceramic sensor. The sensor is arranged in a symmetrical configuration, with one sensor installed on each side of the sample. An acoustic coupling agent is coated between the sensor and the sample surface. A signal trigger threshold is set, and data recording is only initiated when the signal amplitude exceeds the trigger threshold. Stress-strain data is collected synchronously, and the correlation between acoustic emission signals and mechanical responses is realized through timestamp alignment technology. The signal conditioning unit is used to improve the signal-to-noise ratio of weak signals using a low-noise preamplifier; A digital acquisition unit is used to convert analog signals into digital signals.

Images