Intelligent identification method and system for filling body failure stage based on acoustic emission signals

Abstract

The present application relates to the technical field of filling body structure health monitoring, and discloses a filling body damage stage intelligent identification method and system based on acoustic emission signals, which comprises the following steps: S1, collecting acoustic emission original signals in the stress process of the filling body; S2, preprocessing the acoustic emission original signals to obtain a standardized acoustic emission signal sequence; S3, based on the standardized acoustic emission signal sequence, performing self-adaptive stage division through a cumulative signal intensity curve to divide the damage process into several stages; S4, extracting multi-scale features from the acoustic emission signals after stage division, and screening and standardizing the multi-scale features; S5, constructing a deep neural network model based on multi-scale convolution and Transformer fusion, training the deep neural network model by using the screened multi-scale features; and using the trained deep neural network model to identify and classify the damage stages of the filling body.

Description

Technical Field

[0001] This invention relates to the field of infill structure health monitoring technology, specifically to an intelligent identification method and system for infill structure failure stages based on acoustic emission signals. Background Technology

[0002] As the core support structure in backfill mining, the backfill body is formed by the solidification 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 prone to nonlinear evolution processes of microcrack initiation, propagation, and ultimately macroscopic instability and failure under the high ground stress and complex environment of deep mining. Acoustic emission, as a dynamic nondestructive testing technology, can capture transient elastic waves released from damage and rupture within the backfill body in real time and is widely used to monitor its stability. In current technologies, researchers mostly use parameter-based analysis methods, such as using parameters like ring count, energy, and amplitude to calculate b-values, Ib-values, or RA-AF correlation values ​​to assess the damage state. However, such methods mainly rely on empirical formulas and statistical laws, often requiring preset fixed empirical thresholds to classify the failure stages. Due to the diverse composition of mine backfill and the complex stress conditions, fixed thresholds are difficult to adapt to the damage evolution patterns under different cementation strengths or loading rates. As a result, in practical engineering applications, the judgment of the critical failure point is often delayed or misjudged, which cannot meet the needs of high-precision real-time monitoring.

[0003] With the development of artificial intelligence technology, acoustic emission signal recognition methods based on machine learning have gradually become a research hotspot. In existing literature, algorithms such as Support Vector Machine (SVM), Random Forest, and Backpropagation (BP) neural networks have been applied to the classification and regression prediction of acoustic emission signals. However, these traditional machine learning models heavily rely on manually designed statistical features, making feature extraction cumbersome and highly subjective, and failing to fully extract the deep temporal information contained in acoustic emission signals. Although deep learning models such as Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM) have been introduced into this field in recent years, single-scale convolutional kernels struggle to simultaneously capture transient changes (high-frequency details) and long-term trends (low-frequency evolution) in signals. Furthermore, LSTM suffers from gradient vanishing and low computational efficiency when processing extremely long sequences. In addition, existing supervised learning methods typically use simple stress levels for data labeling, lacking a scientific adaptive stage division mechanism, resulting in a large amount of label noise in the training samples, severely limiting the model's generalization ability and recognition accuracy. Therefore, there is an urgent need for an intelligent recognition method for infill body damage stages that can adaptively divide damage stages and integrate multi-scale local features with global long-term dependencies. Summary of the Invention

[0004] To address the technical problems mentioned above, this invention uses deep learning technology to analyze the temporal characteristics of acoustic emission signals from infill bodies, enabling adaptive stage division and high-precision intelligent identification of the infill body failure process, thereby providing a basis for structural health monitoring, reinforcement, and maintenance.

[0005] To achieve the above objectives, the present invention provides an intelligent identification method for the failure stage of infill bodies based on acoustic emission signals, comprising the following steps:

[0006] S1. Collect the original acoustic emission signals during the stress process of the filling body;

[0007] S2. Preprocess the raw acoustic emission signal to obtain a standardized acoustic emission signal sequence;

[0008] S3. Based on the standardized acoustic emission signal sequence, the destruction process is divided into several stages by adaptive stage division through the cumulative signal intensity curve.

[0009] S4. Extract multi-scale features from the acoustic emission signal after the segmentation, and filter and standardize the multi-scale features.

[0010] S5. Construct a deep neural network model based on the fusion of multi-scale convolution and Transformer, train the deep neural network model using the selected multi-scale features, and use the trained deep neural network model to identify and classify the destruction stages of the filling.

[0011] Preferably, S1 includes: a monitoring system consisting of an acoustic emission sensor, a preamplifier, and a data acquisition system, which collects acoustic emission signals of the filling sample throughout the entire stress process;

[0012] When acquiring acoustic emission signals, a signal threshold is set to eliminate background noise, and the number of acoustic emission events is recorded. Before acquiring acoustic emission signals, a standard pencil-breaking test is performed to verify the accuracy and sensitivity of the acoustic emission system.

[0013] Preferably, S2 includes:

[0014] Based on the original acoustic emission signal, clean and format standardize it to obtain data in a uniform format.

[0015] Based on the unified format data, missing values ​​and outliers are processed to generate intermediate sequence data;

[0016] Based on the intermediate sequence data, a basic quality check is performed to output the standardized acoustic emission signal sequence.

[0017] Preferably, S3 includes:

[0018] Based on the standardized acoustic emission signal sequence, target parameters are selected and the cumulative summation sequence of target parameters is calculated to obtain the cumulative signal intensity curve;

[0019] Based on the accumulated signal strength curve, normalization processing is performed to obtain a normalized accumulated signal strength curve;

[0020] Based on the normalized cumulative signal intensity curve, combined with the preset energy threshold and time constraints, the boundary points between multiple destruction stages are determined, thereby completing the stage division.

[0021] Preferably, S4 includes:

[0022] Based on the acoustic emission signals after the stages are divided, RA values ​​and AF values ​​are constructed to obtain the basic feature set;

[0023] Based on the aforementioned basic feature set, multi-scale sliding window statistical analysis is performed on energy, amplitude, RA value, and AF value to obtain a statistical feature set including mean, standard deviation, and first-order difference slope.

[0024] Based on the feature pool composed of the basic feature set and the statistical feature set, feature filtering and standardization are performed to obtain an optimized feature set;

[0025] Based on the optimized feature set, three-dimensional tensor samples are generated by segmentation using a sliding window.

[0026] Preferably, S5 includes: constructing a network structure that includes an input layer, a multi-scale convolutional module, an SE attention mechanism, a Transformer encoder, and a classifier;

[0027] The multi-scale convolution module includes several parallel one-dimensional convolutional layers with different kernel sizes;

[0028] The Transformer encoder includes a multi-layer, multi-head self-attention mechanism and a feedforward network.

[0029] Preferred methods for training deep neural network models include:

[0030] Focal Loss was used as the loss function, and the focusing parameter and label smoothing parameter were set.

[0031] Use the AdamW optimizer to train the model, setting the initial learning rate, weight decay coefficient, and batch size;

[0032] During training, adjust the model parameters based on the performance on the validation set until the model converges.

[0033] The present invention also provides an intelligent identification system for the failure stage of filling materials based on acoustic emission signals. The system is used to implement the above method and includes: a data acquisition module, a processing module, a division module, an extraction module, and a construction module.

[0034] The acquisition module is used to acquire the original acoustic emission signals during the stress process of the filling body;

[0035] The processing module is used to preprocess the raw acoustic emission signal to obtain a standardized acoustic emission signal sequence;

[0036] The segmentation module is used to perform adaptive stage segmentation based on a standardized acoustic emission signal sequence and the cumulative signal intensity curve, dividing the destruction process into several stages.

[0037] The extraction module is used to extract multi-scale features from the acoustic emission signal after the segmentation, and to filter and standardize the multi-scale features.

[0038] The building module is used to construct a deep neural network model based on the fusion of multi-scale convolution and Transformer, train the deep neural network model using the selected multi-scale features, and use the trained deep neural network model to identify and classify the destruction stages of the filling.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] This invention effectively overcomes the limitations of traditional methods that rely on fixed empirical thresholds through an adaptive intelligent stage division mechanism. It can dynamically adapt to the damage evolution process of filling materials under different ratios and loading conditions, thereby significantly improving the accuracy and timeliness of damage stage identification. Attached Figure Description

[0041] 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.

[0042] Figure 1 This is a flowchart illustrating the overall method of an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the intelligent stage division based on CSS curves in an embodiment of the present invention;

[0044] Figure 3 This is a flowchart illustrating the construction process of multi-scale feature engineering according to an embodiment of the present invention;

[0045] Figure 4This is a diagram of the multi-scale convolutional Transformer model architecture according to an embodiment of the present invention;

[0046] Figure 5 This is a performance curve of the model training process in an embodiment of the present invention. Detailed Implementation

[0047] 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.

[0048] 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.

[0049] Example 1:

[0050] like Figure 1 The diagram shown is a schematic representation of the method flow in this embodiment, and the steps include:

[0051] S1. Collect the original acoustic emission signals during the stress process of the filling body.

[0052] The acoustic emission system in this embodiment includes, but is not limited to, an acoustic emission sensor, a preamplifier, a data acquisition system, and signal processing software.

[0053] The raw acoustic emission (AE) signals of the filling specimens were collected throughout the entire stress process. A DS5 AE system, mainly composed of sensors, amplifiers, a data acquisition system, and a personal computer, was used for AE monitoring. The sensor model was RS-2A, with a frequency range of 50-400kHz. Two AE sensors were then fixed to both sides of the specimen, using Vaseline as a coupling agent. An 18mV signal threshold was set to eliminate background noise, and the number of AE events was recorded. Simultaneously, the preamplifier and main amplifier gains were both 40dB, and the sampling frequency was 10MHz. The peak recognition time, impact recognition time, and impact lock-in time were set to 100µs, 200µs, and 300µs, respectively. The sampling frequency of the data acquisition system was also 10MHz. Interpolation was used to obtain the time consistency between the AE acquisition system and the electro-hydraulic servo universal testing machine. A standard pencil break test was performed before each compression test to verify the accuracy and sensitivity of the acoustic emission.

[0054] S2. Preprocess the raw acoustic emission signal to obtain a standardized acoustic emission signal sequence.

[0055] First, the raw data undergoes comprehensive cleaning and format standardization. For non-numeric characters in the Excel data, robust conversion using regular expressions is employed to automatically remove special symbols such as "<", ">", "=", and "±". Percentage values ​​are converted to decimal form, and the average of range values ​​is identified and calculated. Simultaneously, data in scientific notation format is parsed. Next, missing and outlier values ​​are handled: completely empty columns are removed, positive and negative infinity values ​​are replaced with invalid values, and the numerical distribution of each column is examined. Missing values ​​are filled using the median to maintain robust data distribution. Finally, basic data quality checks are performed to ensure the data length meets minimum analytical requirements, verify the peak intensity of key signal columns, and remove invalid dead signals or background noise, thereby constructing a standardized numerical sequence.

[0056] S3. Based on the standardized acoustic emission signal sequence, the destruction process is divided into several stages by adaptive stage division through the cumulative signal intensity curve.

[0057] Adaptive intelligent stage division is performed based on the cumulative signal strength (CSS) curve characteristics. The division logic is as follows: Figure 2 As shown, the failure process of the filling material is divided into an initial stage, a development stage, and a failure stage. The intelligent stage division based on the CSS curve specifically includes: prioritizing the selection of acoustic emission energy parameters; if these are missing, selecting amplitude parameters; calculating their cumulative summation sequence to obtain the CSS curve; and normalizing the CSS curve to the [0, 1] interval using the following formula:

[0058] ,

[0059] in, This represents the normalized cumulative signal strength. This is the original accumulated signal intensity sequence. and These are the maximum and minimum values ​​of the sequence, respectively.

[0060] Determine the stage boundaries based on the time index and energy percentage:

[0061] (1) Initial stage (Stage 0): Directly defined as the first 10% of the data sequence;

[0062] (2) Development Stage (Stage 1): From the end of the initial stage to the beginning of the destruction stage;

[0063] (3) Destruction Stage (Stage 2): The starting point is defined as the time point when the normalized CSS curve first exceeds 0.8, and this time point is subject to constraints and must not be earlier than 60% of the total data length.

[0064] S4. Extract multi-scale features from the acoustic emission signals after segmentation, and then filter and standardize the multi-scale features.

[0065] Feature extraction and filtering process as follows Figure 3 As shown, firstly, physical correlation features are constructed based on the acoustic emission mechanism, and the RA value (ratio of rise time to amplitude) and AF value (ratio of ring count to duration) are calculated, with a small amount introduced during the calculation to prevent the denominator from being zero. Then, a multi-scale sliding window technique is introduced, with window sizes of 10 and 30, to perform rolling statistical analysis on energy, amplitude, RA value, and AF value, respectively, calculating the mean, standard deviation, and first-order difference slope within each window to capture transient changes and short-term trends of the signal. On this basis, a high-dimensional feature pool containing original features and statistical features is constructed, and the correlation between each feature and the damage stage label is evaluated using the SelectKBest algorithm combined with the F-value of analysis of variance. The 60 features with the highest contribution are automatically selected, and RobustScaler is used to robustly standardize the features using the median and interquartile range. Finally, the time series is divided into three-dimensional tensor samples with a length of 64 and a step size of 4 using the sliding window method, which are used as inputs to the deep learning model.

[0066] A multi-scale feature engineering system is constructed to extract the temporal and engineering parameter features of acoustic emission signals, and feature selection is performed. Multi-scale feature engineering includes:

[0067] The RA value feature is constructed, and the calculation formula is as follows:

[0068] ,

[0069] In the formula, RT represents the rise time of the acoustic emission signal, and A AE This indicates the amplitude of the acoustic emission signal.

[0070] The AF value feature is constructed, and the calculation formula is as follows:

[0071] ,

[0072] In the formula, C AE The ring count represents the number of rings in the acoustic emission signal, and DUR represents the duration of the acoustic emission signal.

[0073] Multi-window sliding statistical analysis was performed on the energy, amplitude, RA value, and AF value features, with window sizes set to 10 and 30. The SelectKBest algorithm was used to select the 60 features with the highest contribution from the constructed feature pool as model input.

[0074] S5. Construct a deep neural network model based on the fusion of multi-scale convolution and Transformer, train the deep neural network model using the selected multi-scale features, and use the trained deep neural network model to identify and classify the destruction stages of the filling.

[0075] Construct a deep neural network model based on the fusion of multi-scale convolution and Transformer, with the model architecture as follows: Figure 4 As shown, the deep neural network model consists of an input layer, a multi-scale convolutional module, a SE attention mechanism, a Transformer encoder, and a classifier connected in sequence. The input sequence window size of the input layer is 64. The multi-scale convolutional module contains four parallel one-dimensional convolutional branches with kernel sizes of [3, 7, 15, 1], each branch outputting 32 channels, for a total of 128 channels. The reduction ratio of the SE attention mechanism is set to 8. The Transformer encoder contains a 3-layer stacked structure, with 4 heads for the multi-head self-attention mechanism in each layer, a model dimension of 128, a feedforward network dimension of 512, and a Dropout ratio of 0.4.

[0076] According to the intelligent identification method for the destruction stage of acoustic emission signals in filled bodies based on multi-scale convolutional Transformer provided in this embodiment, the model training adopts Focal Loss loss function to solve the class imbalance problem, and the performance change trend during the training process is as follows. Figure 5 As shown.

[0077] The focusing parameter gamma was set to 1.5, and the label smoothing parameter was set to 0.1; the optimizer used was AdamW, with an initial learning rate of 1×10⁻⁶. -4 The weight decay coefficient is 1×10 -2 The batch size is 128.

[0078] The intelligent identification method for the destruction stage of acoustic emission signals of filling bodies based on multi-scale convolutional Transformer provided in this embodiment is applicable to the stability monitoring and dynamic disaster early warning of filling bodies in mining goaf areas, tailings dam bodies, and rock-fill body composite structures in deep underground engineering.

[0079] Experiments show that the method achieves a recognition accuracy of 99.53%, precision of 98.07%, recall of 99.75%, and F1 score of 98.90% on the validation set. The training process is stable, rapidly improving from 94.38% accuracy in the first epoch to the optimal 99.53%.

[0080] The above results demonstrate that the multi-scale convolutional Transformer-based method proposed in this invention can effectively identify the failure stages of the filling material, exhibiting extremely high accuracy and reliability.

[0081] Example 2:

[0082] This embodiment also provides an intelligent identification system for the failure stages of infill bodies based on acoustic emission signals, including: an acquisition module, a processing module, a segmentation module, an extraction module, and a construction module; the acquisition module is used to acquire the original acoustic emission signals during the stress process of the infill body; the processing module is used to preprocess the original acoustic emission signals to obtain a standardized acoustic emission signal sequence; the segmentation module is used to adaptively segment the failure process into several stages based on the standardized acoustic emission signal sequence and the cumulative signal intensity curve; the extraction module is used to extract multi-scale features from the acoustic emission signals after segmentation, and to filter and standardize the multi-scale features; the construction module is used to construct a deep neural network model based on the fusion of multi-scale convolution and Transformer, train the deep neural network model using the filtered multi-scale features, and use the trained deep neural network model to identify and classify the failure stages of the infill body.

[0083] 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 the failure stage of infill bodies based on acoustic emission signals, characterized in that, Includes the following steps: S1. Collect the original acoustic emission signals during the stress process of the filling body; S2. Preprocess the raw acoustic emission signal to obtain a standardized acoustic emission signal sequence; S3. Based on the standardized acoustic emission signal sequence, adaptive stage division is performed using the cumulative signal intensity curve to divide the destruction process into several stages; S3 includes: selecting target parameters and calculating the cumulative summation sequence of the target parameters according to the standardized acoustic emission signal sequence to obtain the cumulative signal intensity curve; performing normalization processing on the cumulative signal intensity curve to obtain the normalized cumulative signal intensity curve; and determining the boundary points between multiple destruction stages based on the normalized cumulative signal intensity curve and a preset energy threshold and time constraint conditions, thereby completing the stage division. S4. Extract multi-scale features from the acoustic emission signal after segmentation, and filter and standardize the multi-scale features; S4 includes: constructing RA values ​​and AF values ​​based on the acoustic emission signal after segmentation to obtain a basic feature set; performing multi-scale sliding window statistical analysis on energy, amplitude, RA values, and AF values ​​based on the basic feature set to obtain a statistical feature set containing mean, standard deviation, and first-order difference slope; performing feature filtering and standardization processing on the feature pool composed of the basic feature set and the statistical feature set to obtain an optimized feature set; and generating three-dimensional tensor samples by segmenting the optimized feature set using a sliding window. S5. Construct a deep neural network model based on the fusion of multi-scale convolution and Transformer, train the deep neural network model using the selected multi-scale features, and use the trained deep neural network model to identify and classify the destruction stages of the filling.

2. The intelligent identification method for the failure stage of infill bodies based on acoustic emission signals according to claim 1, characterized in that, S1 includes: The monitoring system, consisting of an acoustic emission sensor, a preamplifier, and a data acquisition system, collects acoustic emission signals of the filling sample throughout the entire stress process. When acquiring acoustic emission signals, a signal threshold is set to eliminate background noise, and the number of acoustic emission events is recorded; Before acquiring acoustic emission signals, a standard pencil-breaking test was conducted to verify the accuracy and sensitivity of the acoustic emission system.

3. The intelligent identification method for the failure stage of infill bodies based on acoustic emission signals according to claim 1, characterized in that, S2 include: Based on the original acoustic emission signal, clean and format standardize it to obtain data in a uniform format. Based on the unified format data, missing values ​​and outliers are processed to generate intermediate sequence data; Based on the intermediate sequence data, a basic quality check is performed to output the standardized acoustic emission signal sequence.

4. The intelligent identification method for the failure stage of infill bodies based on acoustic emission signals according to claim 1, characterized in that, S5 includes: constructing a network structure that includes an input layer, a multi-scale convolutional module, an SE attention mechanism, a Transformer encoder, and a classifier; The multi-scale convolution module includes several parallel one-dimensional convolutional layers with different kernel sizes; The Transformer encoder includes a multi-layer, multi-head self-attention mechanism and a feedforward network.

5. The intelligent identification method for the failure stage of infill bodies based on acoustic emission signals according to claim 4, characterized in that, Training methods for deep neural network models include: Focal Loss was used as the loss function, and the focusing parameter and label smoothing parameter were set. Use the AdamW optimizer to train the model, setting the initial learning rate, weight decay coefficient, and batch size; During training, adjust the model parameters based on the performance on the validation set until the model converges.

6. An intelligent identification system for the failure stage of infill bodies based on acoustic emission signals, the system being used to implement the method described in any one of claims 1-5, characterized in that, include: The module includes a data acquisition module, a processing module, a partitioning module, an extraction module, and a construction module. The acquisition module is used to acquire the original acoustic emission signals during the stress process of the filling body; The processing module is used to preprocess the raw acoustic emission signal to obtain a standardized acoustic emission signal sequence; The segmentation module is used to perform adaptive stage segmentation based on a standardized acoustic emission signal sequence and the cumulative signal intensity curve, dividing the destruction process into several stages. The extraction module is used to extract multi-scale features from the acoustic emission signal after the segmentation, and to filter and standardize the multi-scale features. The building module is used to construct a deep neural network model based on the fusion of multi-scale convolution and Transformer, train the deep neural network model using the selected multi-scale features, and use the trained deep neural network model to identify and classify the destruction stages of the filling.

Images