A solid-liquid mixed system concentration prediction method and system based on acoustic emission technology

Abstract

The application provides a solid-liquid mixing system concentration prediction method and system based on acoustic emission technology, and relates to the technical field of acoustic detection.The application collects acoustic emission signals in the mixing process of solid particles and liquid stirring, pre-processes the obtained acoustic emission signals, extracts various modal features of the signals, including time-frequency domain features, wavelet transform features, empirical mode decomposition features and features based on a physical model, selects high-quality signal samples through a classifier, uses a residual network integrated model with an attention mechanism to predict the concentration of the selected signal features, and outputs a prediction result.The application realizes high-precision prediction of the concentration of a solid-liquid mixing system through multi-domain feature fusion and a deep learning model, improves the reliability of the prediction result through uncertainty quantification, further improves the model performance and adaptability by using a hierarchical optimization strategy, and can be widely applied to the monitoring and control of the solid-liquid mixing process in industrial production.

Description

Technical Field

[0001] This invention relates to the field of acoustic detection technology, specifically to a method and system for predicting the concentration of a solid-liquid mixture based on acoustic emission technology. Background Technology

[0002] In industrial production, the concentration of solid particles in a solid-liquid mixture is a critical parameter that directly affects product quality and production efficiency. Traditional concentration detection methods include specific gravity methods, refractive index methods, and ultrasonic methods, but these methods often suffer from problems such as slow response speed, low accuracy, susceptibility to environmental interference, complex theoretical models, and the need for invasive installation.

[0003] Acoustic emission (AE) technology analyzes material properties by detecting stress waves released when materials are subjected to stress or undergo changes in state. It offers advantages such as non-invasiveness, high sensitivity, and good real-time performance, showing promising application prospects in solid-liquid mixture concentration detection. However, the acoustic emission signals generated during solid-liquid mixing and stirring are nonlinear and non-stationary, and are affected by various factors such as rotational speed and particle characteristics. How to efficiently screen and extract effective features from these complex signals to achieve high-precision prediction of solid-liquid mixture concentration remains a core challenge that urgently needs to be overcome in this technological field.

[0004] Concentration prediction methods based on acoustic emission signals typically employ single-domain features or simple machine learning models, making it difficult to fully capture the complex information within the signal, thus requiring improvements in prediction accuracy and robustness. Furthermore, most methods only provide single-point prediction results, lacking quantification of prediction uncertainty, which fails to meet the reliability requirements of industrial applications. In existing technologies, solid-liquid mixture concentration detection often relies on invasive sensors, which are susceptible to interference from the mixing process and have high maintenance costs; some non-invasive methods rely only on single-domain features, making it difficult to comprehensively capture the complex information of concentration changes, thus limiting prediction accuracy; simultaneously, traditional models often only output a single predicted value, lacking quantitative analysis of prediction reliability, making it difficult to meet the high-precision, high-reliability decision-making needs of industrial scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to propose a method for predicting the concentration of solid-liquid mixtures based on acoustic emission technology. By combining multi-domain feature fusion with a deep learning model, high-precision prediction of the concentration of solid-liquid mixtures is achieved, and the reliability of the results is improved through uncertainty quantification. The core innovation lies in employing a hierarchical intelligent optimization strategy. This strategy selects appropriate optimization algorithms for different stages, using the minimization of the root mean square error as the objective function. It automatically optimizes classifier parameters, deep model hyperparameters, feature subsets, and signal filtering thresholds, significantly improving model performance and engineering adaptability, and providing technical support for real-time monitoring and control of industrial mixing processes.

[0006] This invention provides a method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology, comprising:

[0007] Step 1: Collect multiple acoustic emission signals during the mixing process of solid particles and liquid, and perform Butterworth bandpass filtering and normalization on each acoustic emission signal to obtain the preprocessed signal;

[0008] Step 2: For each preprocessed signal, obtain the features based on the physical model, extract features from the preprocessed signal, and fuse them with the features based on the physical model to obtain a fused feature vector;

[0009] Step 3: Input the fused feature vector into the concentration range prediction classifier to obtain the initial concentration range;

[0010] Step 4: Based on the initial concentration range, select the corresponding quality classifier, input the fused feature vector into the quality classifier, and obtain the predicted quality label corresponding to the preprocessed signal, as well as the probability of belonging to the predicted quality label;

[0011] Step 5: Based on the predicted quality label and the probability of belonging to that predicted quality label, filter all fused feature vectors to obtain multiple target fused feature vectors;

[0012] Step 6: Use multiple attention residual network sub-models to process the fusion feature vector of each target to obtain the target concentration and target uncertainty parameters.

[0013] Optionally, step 2 specifically includes:

[0014] The preprocessed signal is subjected to time-domain feature extraction to obtain time-domain features. The preprocessed signal is subjected to fast Fourier transform to obtain the spectrum. The spectrum is then subjected to feature extraction to obtain frequency-domain features.

[0015] The preprocessed signal is decomposed using the db2 wavelet basis to obtain wavelet coefficients at different scales. The energy at each scale is then calculated based on these wavelet coefficients using the following formula:

[0016] ;

[0017] in, For the energy at the j-th scale, For the wavelet coefficients at the j-th scale;

[0018] The total wavelet energy is obtained by summing the energy at all scales. The ratio of the energy at each scale to the total wavelet energy is calculated to obtain the energy fraction. The energy fractions at the preset scales are used to form the wavelet transform features.

[0019] The preprocessed signal is decomposed by empirical mode decomposition to obtain multiple intrinsic mode function components. The first three intrinsic mode function components are obtained. For each of the first three intrinsic mode function components, the decomposition characteristics of the intrinsic mode function component are calculated, including the mean, standard deviation, maximum value, minimum value, peak-to-peak value, and absolute value of the mean. The decomposition characteristics of the first three intrinsic mode function components constitute the empirical mode decomposition characteristics.

[0020] The fused feature vector is obtained by fusing time-domain features, frequency-domain features, wavelet transform features, empirical mode decomposition features, and physical model-based features.

[0021] Optionally, the time-domain features include mean, standard deviation, root mean square (RMS), peak-to-peak value, kurtosis, skewness, zero-crossing rate, absolute maximum value, waveform factor, peak factor, and impulse factor, wherein the absolute maximum value is the maximum amplitude in the preprocessed signal, the waveform factor is the ratio of RMS to the absolute value of the mean, the peak factor is the ratio of the absolute maximum value to RMS, and the impulse factor is the ratio of the absolute maximum value to the absolute value of the mean.

[0022] Optionally, the frequency domain features include: spectral centroid, spectral bandwidth, total energy, spectral roll-off point, and spectral flatness;

[0023] The spectral centroid is calculated using the following formula:

[0024] ;

[0025] Where f i Let X(f) be the frequency. i ) represents the Fourier coefficients for the corresponding frequencies;

[0026] The energy is the square of the spectral amplitude corresponding to the frequency; the spectral bandwidth is the weighted square of the difference between the frequency and the centroid of the spectrum; the total energy is the sum of the squares of the spectral amplitudes; the spectral roll-off point is the frequency in the spectrum where the energy of the frequency accumulates to 85% of the total energy from the lowest frequency to the highest frequency; and the spectral flatness is the ratio of the geometric mean to the arithmetic mean of the spectral amplitudes in the spectrum.

[0027] Optionally, the physical model-based features are obtained in the following ways:

[0028] Experiments were conducted on the particle collision process in a solid-liquid mixture, and the following parameters were collected: rotational frequency f, concentration correction factor K, particle mass m, tangential velocity V, frictional work W, and conversion efficiency. Sensor area S and rotational speed value;

[0029] Calculate the number of collisions per unit time between particles and the container wall in a solid-liquid mixture. Specifically, this is achieved through the following formula:

[0030] ;

[0031] Calculate the energy range of particle collisions with the container wall in a solid-liquid mixture. Specifically, this is achieved through the following formula:

[0032] ;

[0033] Calculate the acoustic emission energy captured by the sensor Specifically, this is achieved through the following formula:

[0034] ;

[0035] The mean, median, skewness, and kurtosis of the preprocessed signal are obtained to acquire the signal's statistical characteristics.

[0036] Based on the sum of the squares of the rotational speed values The speed-related characteristics are calculated and expressed as follows:

[0037] ;

[0038] The , , The features based on the physical model consist of signal statistical characteristics and speed-related characteristics.

[0039] Optionally, the concentration range prediction classifier in step 3 is trained in the following way:

[0040] The historical fusion feature vector is obtained as the first input sample, and the concentration range label corresponding to the historical fusion feature vector is obtained as the first output sample. The first input sample and the first output sample form the first training sample, and multiple first training samples form the first training sample set.

[0041] Multiple classifiers are obtained, including SVM, KNN, Random Forest, and XGBoost. Each classifier is trained based on the first training sample set to obtain the first trained classifier. The classifier with the highest accuracy among all the first classifiers is selected as the concentration range prediction classifier.

[0042] Optionally, the quality classifier in step 4 can be trained in the following way:

[0043] A historical fusion feature vector is obtained as a second input sample, and the corresponding quality label is obtained as a second output sample. When the preprocessed signal corresponding to the historical fusion feature vector meets the abnormal conditions, the quality label is an invalid signal; when the preprocessed signal corresponding to the historical fusion feature vector does not meet the abnormal conditions, the quality label is a valid signal. The abnormal conditions include at least one of the following: signal-to-noise ratio is lower than a preset threshold, signal amplitude is lower than a preset threshold, the proportion of abnormal noise in the spectrum exceeds a preset ratio, sensor output saturation, or missing signal data. The second input sample and the second output sample constitute the second training sample, and multiple second training samples constitute the second training sample set.

[0044] The second training sample set is divided into multiple sub-training sets according to the concentration range of the historical fusion feature vectors;

[0045] Multiple classifiers are obtained, including SVM, KNN, Random Forest, and XGBoost. For each sub-training set, each classifier is trained using the sub-training set to obtain a trained second classifier. The classifier with the highest accuracy among all the second classifiers is selected as the quality classifier for the concentration range corresponding to the sub-training set, thus obtaining the quality classifier for each concentration range.

[0046] Optionally, step 5 includes:

[0047] The fusion feature vectors whose predicted quality labels match the true quality labels corresponding to the fusion feature vectors and whose probability is greater than a preset threshold are obtained and used as target fusion feature vectors. All target fusion feature vectors form the optimal feature subset.

[0048] Optionally, step 6 includes:

[0049] By simultaneously inputting the target fused feature vector into multiple attention residual network sub-models, multiple mean predicted concentrations and multiple intra-model uncertainty parameters are obtained:

[0050] Among them, the attention residual network sub-models have the same structure but different parameters;

[0051] In the attention residual network sub-model, the target fusion feature vector passes through multiple sequential feature extraction modules and then through the output layer to obtain the mean predicted concentration and the uncertainty parameters within the model.

[0052] The feature extraction module includes a residual block and an attention block. In the first feature extraction module, the target fused feature vector is passed sequentially through a linear layer, a batch normalization layer, a ReLU activation function, another linear layer, and another batch normalization layer to obtain transformed features. The transformed features are added to the fused feature vector, and the added features are passed through a ReLU activation function to obtain residual concatenated features. The residual concatenated features are passed sequentially through a linear layer, a ReLU activation function, a linear layer, and a Sigmoid activation function to generate attention weights. The attention weights are multiplied element-wise with the residual concatenated features to obtain weighted features. The weighted features are input to the next feature extraction module until the last feature extraction module outputs high-level features.

[0053] The high-level features are input into the output layer. In the output layer, the high-level features pass through a linear layer to obtain the predicted concentration mean. At the same time, the high-level features pass through a linear layer and a Softplus activation function to obtain the standard deviation, which is used as the uncertainty parameter within the model.

[0054] The arithmetic mean of the predicted concentrations output by all attention residual network sub-models is calculated to obtain the target concentration. The arithmetic mean of the intra-model uncertainty parameters output by all attention residual network sub-models is calculated to obtain the final intra-model uncertainty parameters. The standard deviation of the predicted concentrations output by all attention residual network sub-models is calculated to obtain the inter-model uncertainty parameters. The final intra-model uncertainty parameters and the inter-model uncertainty parameters constitute the target uncertainty parameters.

[0055] A concentration prediction system for solid-liquid mixtures based on acoustic emission technology, comprising:

[0056] The signal acquisition module is used to acquire multiple acoustic emission signals during the solid-liquid mixing process;

[0057] The signal preprocessing module is used to perform Butterworth bandpass filtering and normalization on each acoustic emission signal to obtain the preprocessed signal.

[0058] The feature extraction and fusion module is used to obtain physical model-based features for each preprocessed signal, extract features from the preprocessed signal, and fuse them with the physical model-based features to obtain a fused feature vector.

[0059] The signal filtering module is used to input the fused feature vector into the concentration range prediction classifier to obtain the initial concentration range. Based on the initial concentration range, the corresponding quality classifier is selected, and the fused feature vector is input into the quality classifier to obtain the predicted quality label corresponding to the preprocessed signal and the probability of belonging to the predicted quality label. Based on the predicted quality label and the probability of belonging to the predicted quality label, all fused feature vectors are filtered to obtain multiple target fused feature vectors.

[0060] The prediction model module is used to process the fused feature vector of each target using multiple attention residual network sub-models to obtain the target concentration and target uncertainty parameters.

[0061] The beneficial effects of adopting the above technical solution are as follows:

[0062] 1. Multi-domain feature fusion enables comprehensive information capture. This invention overcomes the limitations of traditional methods that rely on single-domain features. By extracting data-driven features from the time domain, frequency domain, wavelet transform, and EMD, and combining them with physical features based on particle collision theory, a multi-dimensional feature set is formed. Data-driven features capture the dynamic laws of signals, while physical features embed the intrinsic mechanism of solid-liquid mixing. The complementary fusion of these two approaches can comprehensively reflect the acoustic emission characteristics at different concentrations, laying the foundation for high-precision prediction. Experimental verification shows that multi-domain feature fusion improves the prediction accuracy (R²) by 15%~20% compared to single-feature methods.

[0063] 2. Signal quality filtering enhances model robustness. Addressing the issue of acoustic emission signals being susceptible to interference from equipment vibration and environmental noise in industrial settings, this invention employs a method of training a classifier across different concentration ranges to screen signal samples for quality. The differences in signal characteristics at different concentrations are utilized to optimize classification accuracy, eliminating 10%–15% of low-quality samples due to noise contamination, sensor anomalies, etc., thus reducing invalid signal interference. In practical applications, the model trained using the filtered samples shows a 10%–12% reduction in prediction error (RMSE) and a significant improvement in stability.

[0064] 3. High-precision prediction achieved through an attention-based residual network ensemble model. The ensemble model proposed in this invention addresses the vanishing gradient problem in deep networks through residual connections, ensuring effective feature transfer; it dynamically focuses on key features (such as high-concentration focusing collision intensity and low-concentration focusing spectral features) using an attention mechanism, enhancing sensitivity to concentration changes; and it reduces the bias and variance of individual models through multi-sub-model ensemble. Comparative experiments show that this model improves the prediction accuracy (MAE) by 8%–15% compared to traditional neural networks (such as CNN and LSTM), and exhibits stronger adaptability to different mixed conditions (such as different rotational speeds and particle diameters).

[0065] 4. Uncertainty Quantification Enhances Decision-Making Reliability. This invention not only outputs concentration predictions but also quantifies the reliability of the prediction results by calculating the uncertainties within the model (single-model fluctuations) and between models (sub-model differences). In industrial scenarios, this quantification can assist operators in judging the credibility of the prediction (e.g., manual verification is required when uncertainty is too high), avoiding decision-making risks caused by a single prediction value.

[0066] 5. Layered Intelligent Optimization Strategy Enhances Engineering Practicality. This invention, tailored to the characteristics of classifiers and deep models, employs a layered strategy of grid search and intelligent optimization to automatically optimize model hyperparameters, feature subsets, and filtering thresholds. This addresses the problems of low efficiency and poor adaptability associated with traditional methods involving manual parameter tuning. Experimental results demonstrate that after layered optimization, the model's RMSE is further reduced, and it can automatically adapt to mixed systems in different industrial scenarios without manual intervention.

[0067] 6. Non-invasive detection expands industrial application scenarios. This invention achieves concentration detection based on acoustic emission signals. The sensor does not need to directly contact the mixed materials, avoiding the wear and contamination problems of traditional invasive sensors and reducing maintenance costs. The system response time is ≤1 second, supports real-time acquisition and prediction, and can be integrated into industrial control systems to realize online monitoring and closed-loop control of the mixing process. It is suitable for scenarios such as stirred tanks and pipeline mixing in chemical, pharmaceutical, and food industries. Attached Figure Description

[0068] Figure 1 This is a flowchart illustrating a method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology in an embodiment of the present invention.

[0069] Figure 2 This is a flowchart illustrating another method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology in an embodiment of the present invention.

[0070] Figure 3 This is the concentration prediction result in the embodiment of the present invention. Detailed Implementation

[0071] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0072] Addressing the core challenges in solid-liquid mixture concentration detection—namely, the difficulty in extracting effective information from complex acoustic emission signals and the limited prediction accuracy and lack of reliability quantification in traditional methods—this paper proposes a method and system for predicting solid-liquid mixture concentration that combines high prediction accuracy with uncertainty quantification capabilities. This approach holds significant practical research value and application prospects. Specifically, the system integrates three core components: an acoustic emission sensor, a data acquisition module, and an intelligent analysis unit. During industrial equipment operation, the sensor first captures the acoustic emission signal generated by solid-liquid mixing in real time. After initial signal acquisition and transmission by the data acquisition module, the intelligent analysis unit further extracts multi-domain features of the signal, including time-domain, frequency-domain, and wavelet transform characteristics. Then, based on an optimized algorithm model, these features are analyzed in depth to ultimately achieve accurate prediction of the solid-liquid mixture concentration, providing data support for real-time control of industrial mixing processes.

[0073] To address the problems existing in the prior art, this invention provides a method for predicting the concentration of solid-liquid mixtures based on acoustic emission technology, combined with... Figure 1 and Figure 2 This may include the following steps:

[0074] Step 1: Collect multiple acoustic emission signals during the mixing process of solid particles and liquid, and perform Butterworth bandpass filtering and normalization on each acoustic emission signal to obtain the preprocessed signal;

[0075] Specifically, two high-performance piezoelectric acoustic emission sensors are used and arranged on the sidewalls at different heights along the axial direction of the mixing container to simultaneously acquire acoustic emission signals from different locations. These signals are then connected to a data acquisition card via a signal amplifier to ensure the complete capture of dynamic processes such as particle collisions and liquid disturbances. The acoustic emission signals are expressed as follows: transverse time (seconds) and ordinate amplitude (volts).

[0076] The Butterworth bandpass filter uses a fourth-order filter with a low-frequency cutoff of 10 kHz (to filter out low-frequency ambient noise) and a high-frequency cutoff of 499 kHz (to avoid aliasing interference of sampling frequencies), effectively preserving the main frequency components of acoustic emission generated by particle collisions. The standardization process uses maximum value normalization to compress the signal amplitude to the [-1,1] range, eliminating the influence of signal intensity differences at different acquisition times and providing a uniform scale input for subsequent feature extraction.

[0077] Step 2: For each preprocessed signal, obtain features based on the physical model, extract features from the preprocessed signal, and fuse them with the features based on the physical model to obtain a fused feature vector, including:

[0078] The preprocessed signal is subjected to time-domain feature extraction to obtain time-domain features, including mean (reflecting the overall energy level of the signal), standard deviation (reflecting the degree of signal fluctuation), root mean square (RMS) (characterizing the effective value of the signal), peak-to-peak value (the difference between the maximum and minimum values ​​of the signal, reflecting extreme fluctuations), kurtosis (describing the impulse characteristics of the signal), skewness (reflecting the symmetry of the signal distribution), zero-crossing rate, absolute maximum value, waveform factor, peak factor, and impulse factor. The absolute maximum value is the maximum amplitude in the preprocessed signal, the waveform factor is the ratio of RMS to the absolute value of the mean, the peak factor is the ratio of the absolute maximum value to RMS, and the impulse factor is the ratio of the absolute maximum value to the absolute value of the mean.

[0079] The zero-crossing rate is calculated using the following formula:

[0080] ZCR= ;

[0081] Where N is the number of sampling points, N=1000000, x i Let i be the signal value at the i-th sampling point of the preprocessed signal. For a sign function, when When <0, =1, otherwise 0;

[0082] The formula for calculating the root mean square (RMS) value is as follows: ;

[0083] The preprocessed signal is subjected to a Fast Fourier Transform (FFT) to obtain the spectrum. Feature extraction is then performed on the spectrum to obtain frequency domain features, including: spectral centroid, spectral bandwidth, total energy, spectral roll-off point, and spectral flatness. The spectral centroid is the weighted average of frequency and energy in the spectrum. Energy is the square of the spectral amplitude corresponding to the frequency. The spectral centroid reflects the dominant frequency location where energy is concentrated. The spectral bandwidth is the weighted average of the squares of the differences between frequencies in the spectrum and the spectral centroid, used to describe the energy distribution range. The total energy is the sum of the squares of the spectral amplitudes, characterizing the total signal energy. The spectral roll-off point is the frequency at which the energy of a frequency accumulates to 85% of the total energy from the lowest to the highest frequency in the spectrum, used to reflect the proportion of high-frequency components. The spectral flatness is the ratio of the geometric mean to the arithmetic mean of the spectral amplitudes in the spectrum, used to distinguish noise from periodic signals.

[0084] The formula for calculating the spectral centroid is as follows: , where f i Let X(f) be the frequency. i ) represents the Fourier coefficients for the corresponding frequencies;

[0085] The preprocessed signal is decomposed using the db2 wavelet basis to obtain wavelet coefficients at different scales. The energy at each scale is then calculated based on these wavelet coefficients using the following formula:

[0086] ;

[0087] in, For the energy at the j-th scale, For the wavelet coefficients at the j-th scale;

[0088] The total wavelet energy is obtained by summing the energy at all scales. The ratio of the energy at each scale to the total wavelet energy is calculated to obtain the energy fraction. The energy fractions at preset scales are used to form wavelet transform features. Specifically, the energy fractions at scales d4-d6 (corresponding to the 15.625~62.5 kHz frequency band) are selected. This frequency band is highly matched with the dominant frequency of acoustic emission from particle collisions. A total of 4 wavelet features are obtained, taking into account the local characteristics in both the time and frequency domains.

[0089] The preprocessed signal is decomposed using Empirical Mode Decomposition (EMD) to obtain multiple intrinsic mode function (IMF) components. The first three IMF components are then selected. For each of the first three IMF components, the decomposition features of the IMF component are calculated, including the mean, standard deviation, maximum value, minimum value, peak-to-peak value, and absolute value of the mean. The decomposition features of the first three IMF components constitute the Empirical Mode Decomposition (EMD) features. When there are fewer than three IMF components, the missing features are filled with zeros to ensure uniform feature dimensions and effectively capture the nonlinear and non-stationary characteristics of the signal.

[0090] The physical model-based features are obtained in the following ways:

[0091] Experiments were conducted on the particle collision process in a solid-liquid mixture, and the following parameters were collected: rotational frequency f, concentration correction factor K, particle mass m, tangential velocity V, frictional work W, and conversion efficiency. Sensor area S and rotational speed value;

[0092] Wherein, rotational frequency f is in Hz, f = n / 60, where n is the stirring speed in rpm; concentration correction coefficient K is K = 1 + 0.1 × C, where C is the particle volume concentration in vol%. Particle mass m is in kg; tangential velocity V is in m / s; frictional work W is in J / m²; sensor area S is in m².

[0093] Collision frequency of solid particles with container wall: Calculate the number of collisions between particles and container wall per unit time in a solid-liquid mixture. Specifically, this is achieved through the following formula:

[0094] ;

[0095] Collision intensity range between solid particles and container wall: Calculate the energy range of collisions between particles and container wall in a solid-liquid mixture. Specifically, this is achieved through the following formula:

[0096] ;

[0097] Acoustic emission energy: Calculate the acoustic emission energy captured by the sensor. Specifically, this is achieved through the following formula:

[0098] ;

[0099] The mean, median, skewness, and kurtosis of the preprocessed signal are obtained to acquire the signal's statistical characteristics.

[0100] Based on the sum of the squares of the rotational speed values The speed-related characteristics are calculated and expressed as follows:

[0101] ;

[0102] In the above formula, to avoid the meaninglessness of logarithmic operations when the acoustic emission energy is 0 and the mathematical error of the denominator being 0 when the rotational speed is 0, and to reduce the interference of extreme small values ​​on the results, the numerator and denominator are respectively added with the number 1 when calculating the ratio.

[0103] The , , The features based on the physical model consist of signal statistical characteristics and speed-related characteristics;

[0104] The fused feature vector is obtained by fusing time-domain features, frequency-domain features, wavelet transform features, empirical mode decomposition features, and physical model-based features.

[0105] Step 3: Input the fused feature vector into the concentration range prediction classifier to obtain the initial concentration range;

[0106] The concentration range prediction classifier is trained in the following manner:

[0107] The historical fusion feature vector is obtained as the first input sample, and the concentration range label corresponding to the historical fusion feature vector is obtained as the first output sample. The first input sample and the first output sample form the first training sample, and multiple first training samples form the first training sample set.

[0108] Multiple classifiers are obtained, including SVM, KNN, Random Forest, and XGBoost. Each classifier is trained based on the first training sample set to obtain the first classifier after training. The classifier with the highest accuracy among all the first classifiers is selected as the concentration range prediction classifier.

[0109] Step 4: Based on the initial concentration range, select the corresponding quality classifier, input the fused feature vector into the quality classifier, and obtain the predicted quality label corresponding to the preprocessed signal, as well as the probability of belonging to the predicted quality label;

[0110] The quality classifier is trained in the following manner:

[0111] A historical fusion feature vector is obtained as a second input sample, and the corresponding quality label is obtained as a second output sample. When the preprocessed signal corresponding to the historical fusion feature vector meets the abnormal conditions, the quality label is an invalid signal; when the preprocessed signal corresponding to the historical fusion feature vector does not meet the abnormal conditions, the quality label is a valid signal. The abnormal conditions include at least one of the following: signal-to-noise ratio is lower than a preset threshold, signal amplitude is lower than a preset threshold, the proportion of abnormal noise in the spectrum exceeds a preset ratio, sensor output saturation, or missing signal data. The second input sample and the second output sample constitute the second training sample, and multiple second training samples constitute the second training sample set.

[0112] Based on the concentration range of the historical fusion feature vectors, such as low (0~20%), medium (21~50%), and high (51~100%), the second training sample set is divided into multiple sub-training sets; based on the concentration value, for the samples in each concentration range, SimpleImputer is used to fill in missing features (the strategy is mean); if there are classes with less than 2 samples in a certain concentration range, SMOTE oversampling is used to balance the dataset;

[0113] Multiple classifiers are obtained, including SVM, KNN, Random Forest, and XGBoost. For each sub-training set, each classifier is trained using the sub-training set to obtain a trained second classifier. Among all the second classifiers, the classifier with the highest accuracy is selected as the quality classifier for the concentration range corresponding to the sub-training set, thus obtaining the quality classifier for each concentration range.

[0114] For the four classifiers (SVM, KNN, Random Forest, and XGBoost) in the signal filtering step, grid search is used for parameter optimization to ensure classification accuracy and stability. The optimization targets are the key hyperparameters of each classifier (such as the regularization parameter C and kernel function of SVM, and the weighting method of KNN). The optimization process involves performing grid search separately for each concentration interval, combined with 5-fold hierarchical cross-validation, to select the model with the highest accuracy and corresponding parameters as the optimal classifier for that interval. Its advantage is that it traverses the preset parameter space, ensuring that local optima are found, thus adapting to the differences in signal characteristics across different concentration intervals.

[0115] The optimal classifier and parameters were selected using grid search and 5-fold cross-validation. The grid search parameters were: regularization parameter C (0.1, 1, 10) and kernel function (linear, rbf) for SVM; number of neighbors (3, 5, 7) and weighting method (uniform, distance) for KNN; number of trees (100, 200) and maximum depth (None, 5, 10) for Random Forest; and number of trees (100, 200), learning rate (0.01, 0.1), and maximum depth (3, 5) for XGBoost. Finally, the model with the highest accuracy within each concentration range was selected as the optimal classifier.

[0116] Step 5: Based on the predicted quality label and the probability of belonging to that predicted quality label, filter all fused feature vectors to obtain the target fused feature vector. All target fused feature vectors form the optimal feature subset.

[0117] The fusion feature vectors whose predicted quality labels match the true quality labels corresponding to the fusion feature vectors and whose probability is greater than a preset threshold are obtained and used as target fusion feature vectors. All target fusion feature vectors form the optimal feature subset. The preset threshold is determined by searching within the range of 0.6 to 0.9 using an intelligent optimization algorithm. Experimental verification shows that when the preset threshold is 0.7, more than 85% of low-quality data can be removed while retaining a sufficient sample size.

[0118] Step 6: Use multiple attention residual network sub-models to process the target fusion feature vector to obtain the target concentration and target uncertainty parameters;

[0119] By simultaneously inputting the target fused feature vector into multiple attention residual network sub-models, multiple mean predicted concentrations and multiple intra-model uncertainty parameters are obtained:

[0120] Among them, the attention residual network sub-models have the same structure but different parameters;

[0121] In the attention residual network sub-model, the target fusion feature vector passes through multiple sequential feature extraction modules and then through the output layer to obtain the mean predicted concentration and the uncertainty parameters within the model.

[0122] The feature extraction module includes a residual block and an attention block. In the first feature extraction module, the target fused feature vector is passed sequentially through a linear layer, a batch normalization layer, a ReLU activation function, another linear layer, and another batch normalization layer to obtain transformed features. The transformed features are added to the fused feature vector, and the added features are passed through a ReLU activation function to obtain residual concatenated features. The residual concatenated features are passed sequentially through a linear layer, a ReLU activation function, a linear layer, and a Sigmoid activation function to generate attention weights. The attention weights are multiplied element-wise with the residual concatenated features to obtain weighted features. The weighted features are input to the next feature extraction module until the last feature extraction module outputs high-level features.

[0123] The high-level features are input into the output layer. In the output layer, the high-level features pass through a linear layer to obtain the predicted concentration mean. At the same time, the high-level features pass through a linear layer and a Softplus activation function to obtain the standard deviation, which is used as the uncertainty parameter within the model.

[0124] The dataset was divided into two sets: 80% of the 268 samples were used as the training set (214 sets) and 20% were used as the validation set (54 sets).

[0125] Optimizer: AdamW (lr=1e-4, weight_decay=1e-5);

[0126] Loss function: Negative log-likelihood loss ;

[0127] in: This represents the average loss value for the batch. The total number of samples in a batch; Let be the mean of the predicted concentrations for the i-th sample; Let y be the prediction standard deviation of the i-th sample, i.e., the intra-model uncertainty parameter, used to quantify the uncertainty of the model's prediction for that sample; i Let be the true concentration value of the i-th sample.

[0128] The OneCycleLR strategy is used to schedule the learning rate, with the maximum learning rate set to 1×10⁻³ and the total number of training rounds being 300.

[0129] An early stopping mechanism is introduced, with a patience threshold of 30 rounds and a minimum change threshold of 0.001. The training process stops at the 187th round, and the best model with a validation set loss of 0.023 is obtained.

[0130] For the hyperparameters, feature subsets, and signal filtering threshold of the attention residual network ensemble model, an intelligent optimization algorithm is used for joint optimization. PSO (Particle Swarm Optimization) is used by default, but can be flexibly switched to random search or Bayesian optimization (Optuna). The model hyperparameters are: learning rate (1e-6~5e-4), weight decay (1e-7~1e-3), hidden layer width scaling factor (0.5~2.0), batch size (8~64), number of training epochs (60~180), and number of sub-models (2~6). The feature subset is selected from the fused features based on feature importance scores (selection ratio 0.4~1.0). The signal filtering threshold is the classifier confidence threshold (0.6~0.9). The optimization process first initializes the model by determining the candidate feature pool based on feature importance analysis and setting the search range and constraints for the optimization parameters. Then, for iterative optimization, the intelligent algorithm searches for the optimal parameter combination with RMSE as the objective function, performing model training and performance evaluation in each iteration. Finally, the output includes the optimal hyperparameters, feature subset mask, and filtering threshold, which are used for final model training. PSO is the default algorithm, using collective collaboration among particle swarm optimization to search for the global optimum, resulting in fast convergence and strong robustness. Random search can be used, suitable for rapid validation or scenarios with small datasets, offering simplicity and low computational cost. Bayesian optimization can also be used, guided by a probabilistic model, offering high sample efficiency and suitability for high-dimensional parameter space optimization.

[0131] The target concentration is obtained by calculating the arithmetic mean of the predicted concentrations output by all attention residual network sub-models. The final intra-model uncertainty parameters are then calculated by calculating the arithmetic mean of the intra-model uncertainty parameters output by all attention residual network sub-models. Finally, the inter-model uncertainty parameters are calculated by calculating the standard deviation of the predicted concentrations output by all attention residual network sub-models. The final intra-model uncertainty parameters and the inter-model uncertainty parameters together constitute the target uncertainty parameters.

[0132] Furthermore, by adding or subtracting 1.96 times the target uncertainty parameter from the target mean, the confidence interval is obtained.

[0133] This invention also provides a concentration prediction system for solid-liquid mixtures based on acoustic emission technology, comprising:

[0134] The signal acquisition module is used to acquire multiple acoustic emission signals during the solid-liquid mixing process;

[0135] The signal preprocessing module is used to perform Butterworth bandpass filtering and normalization on each acoustic emission signal to obtain the preprocessed signal.

[0136] The feature extraction and fusion module is used to obtain physical model-based features for each preprocessed signal, extract features from the preprocessed signal, and fuse them with the physical model-based features to obtain a fused feature vector.

[0137] The signal filtering module is used to input the fused feature vector into the concentration range prediction classifier to obtain the initial concentration range. Based on the initial concentration range, the corresponding quality classifier is selected, and the fused feature vector is input into the quality classifier to obtain the predicted quality label corresponding to the preprocessed signal and the probability of belonging to the predicted quality label. Based on the predicted quality label and the probability of belonging to the predicted quality label, all fused feature vectors are filtered to obtain multiple target fused feature vectors.

[0138] The prediction model module is used to process the fused feature vector of each target using multiple attention residual network sub-models to obtain the target concentration and target uncertainty parameters.

[0139] Furthermore, the present invention may also include a visualization module: generating a comparison chart of predicted and true values ​​(with 95% confidence interval error bars), a histogram of prediction error distribution (labeled with mean and standard deviation), a curve of concentration interval error variation, a bar chart of feature importance, and a heatmap of classification confusion matrix;

[0140] Data storage module: Stores raw signals, extracted features, model parameters, prediction results and visualization charts, and supports data backtracking and model retraining.

[0141] Based on the above technical solution, this invention takes a glass bead-water mixture system as the research object and conducts solid-liquid mixing experiments in a 2L stirred tank. The glass beads have a diameter range of 0.4~0.6 mm and a density of 2500 kg / m³, the water density is 1000 kg / m³, the stirring speed range is 0~1500 rpm, and the concentration range is 0~4%. The concentration is predicted using the method and system described in this invention. The specific steps are as follows:

[0142] Signal acquisition utilizes the signal acquisition module of the system described in this invention, comprising two acoustic emission signal acquisition sensors fixed to the axial sidewalls of the stirred tank at different heights (30 mm spacing, 20 mm from the bottom of the tank); a signal amplifier (40 dB gain, bandwidth 1 kHz to 1 MHz); and an acoustic emission signal acquisition device. During the experiment, the data acquisition system is activated to simultaneously acquire dual-channel acoustic emission signals. Each acquisition session lasts 1 second, and 5 sets of samples are acquired for each concentration-speed combination (particle mass concentration 0–4%, with each 0.5% increment representing a gradient; speed 600–1500 rpm, with each 100 rpm increment representing a gradient).

[0143] Signal preprocessing: The system's signal preprocessing module processes the acquired raw signals. Filtering: The filtering unit uses a 4th-order Butterworth bandpass filter with a low-frequency cutoff of 10 kHz and a high-frequency cutoff of 499 kHz. Zero-phase filtering is employed to eliminate low-frequency mechanical vibrations (<10 kHz) and high-frequency electromagnetic interference (>500 kHz) other than particle collisions. Outlier handling: The outlier handling unit detects impulse noise based on the 3σ criterion. For sampling points exceeding [μ-3σ, μ+3σ] (μ is the signal mean, σ is the standard deviation), a moving average of 5 adjacent points is used for correction. Standardization: The standardization unit uses maximum value normalization, calculated using the following formula:

[0144] ;

[0145] Where x is the filtered signal; For the preprocessed signal, the signal amplitude is compressed to the range of [-1, 1].

[0146] The feature extraction and fusion module extracts and fuses time-domain features, frequency-domain features, wavelet transform features, empirical mode decomposition features, and physical model-based features.

[0147] Signal quality filtering: Screening high-quality samples using a signal filtering module.

[0148] Classifier training: The classifier training unit trains a 4-class classifier for each concentration.

[0149] SVM: Parameter range C=[0.1,1,10], kernel function =linear / rbf;

[0150] KNN: n_neighbors=[3,5,7], weights=uniform / distance;

[0151] Random Forest: n_estimators=[100,200], max_depth=[None,5,10];

[0152] XGBoost: n_estimators=[100,200], learning_rate=[0.01,0.1].

[0153] Where C is the regularization parameter, which controls the degree of penalty for misclassified samples; linear / rbf is the kernel function type; n_neighbors is the number of nearest neighbors, i.e., the number of nearest samples referenced during classification; uniform / distance is the weight calculation method, where uniform means that all nearest neighbor samples have the same weight, and distance means that the weight of the nearest neighbor sample is inversely proportional to the distance, and the closer the sample is, the greater its influence on the classification; n_estimators is the number of decision trees, i.e., the total number of decision trees included in the ensemble model; max_depth is the maximum depth of the decision tree, which controls the complexity of a single decision tree; n_estimators is the number of basic decision trees, i.e., the total number of weak classifiers in the ensemble model; and learning_rate is the learning rate (step size), which controls the degree of contribution of each decision tree to the model's prediction results.

[0154] The parameter optimization unit used 5-fold hierarchical cross-validation to select the best model based on accuracy. The best model for the 15%-20% concentration range was XGBoost (n_estimators=200, learning_rate=0.1, max_depth=5), with a cross-validation accuracy of 92.3%.

[0155] Sample screening: The sample screening unit retains samples that are correctly classified and have a prediction probability ≥ 0.7.

[0156] Concentration prediction: Concentration prediction is achieved through the prediction model module.

[0157] Model Construction: The model construction unit generates 5 attention residual network sub-models, each with the following structure:

[0158] Input layer (67-dimensional features, 11+5+4+18+9×2 channel fusion);

[0159] Residual block 1 (128-dimensional) → Attention block 1 → Residual block 2 (64-dimensional) → Attention block 2 → Residual block 3 (32-dimensional) → Attention block 3;

[0160] Output layer: Mean header (32→1), Standard deviation header (32→1+Softplus).

[0161] S52. Model Training: Performed by the training unit, specifically including:

[0162] Dataset split: 80% of the 268 samples are training set (214 groups), and 20% are validation set (54 groups).

[0163] Optimizer: AdamW (lr=1e-4, weight_decay=1e-5);

[0164] Loss function: Negative log-likelihood loss ;

[0165] in: This represents the average loss value for the batch. The total number of samples in a batch; Let be the predicted mean concentration of the i-th sample; Let be the prediction standard deviation of the i-th sample, used to quantify the uncertainty of the model's prediction for that sample; y i Let be the true concentration value of the i-th sample.

[0166] The OneCycleLR strategy is used to schedule the learning rate, with the maximum learning rate set to 1×10⁻³ and the total number of training rounds being 300.

[0167] An early stopping mechanism is introduced, with a patience threshold of 30 rounds and a minimum change threshold of 0.001. The training process stops at the 187th round, and the best model with a validation set loss of 0.023 is obtained.

[0168] S53. The integrated prediction step is executed by the integrated prediction unit, including:

[0169] Calculate the predicted mean, which is the arithmetic mean of the outputs of the five sub-models;

[0170] The uncertainty within the model is calculated as the arithmetic mean of the standard deviations of the outputs of the five sub-models;

[0171] The uncertainty between the models is calculated as the standard deviation of the means of the five sub-model outputs;

[0172] The total uncertainty is calculated based on the intra-model uncertainty and the inter-model uncertainty. Its value is the square root of the sum of the squares of the intra-model uncertainty and the squares of the inter-model uncertainty.

[0173] S6. Result Output and Analysis: The outputs of the result output module and auxiliary module are as follows:

[0174] Performance metrics: MAE=0.32%, RMSE=0.45%, R²=0.987 on the test set (54 groups);

[0175] Uncertainty Coverage: 95% Confidence Interval The coverage rate reached 92.6%;

[0176] Feature importance: The feature importance analysis module shows that wavelet d5 scale energy fraction (weight 12.3%), collision frequency (9.8%), and acoustic emission energy (8.7%) are the most critical features.

[0177] Figure 3The figure shows the concentration prediction results. As can be seen, this invention can achieve high-precision prediction of solid-liquid mixture concentration through multi-domain feature fusion and attention residual network integrated model, and can quantify the uncertainty of prediction to meet the needs of industrial applications.

[0178] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology, characterized in that, include: Step 1: Collect multiple acoustic emission signals during the mixing process of solid particles and liquid, and perform Butterworth bandpass filtering and normalization on each acoustic emission signal to obtain the preprocessed signal; Step 2: For each preprocessed signal, obtain the features based on the physical model, extract features from the preprocessed signal, and fuse them with the features based on the physical model to obtain a fused feature vector; Step 3: Input the fused feature vector into the concentration range prediction classifier to obtain the initial concentration range; Step 4: Based on the initial concentration range, select the corresponding quality classifier, input the fused feature vector into the quality classifier, and obtain the predicted quality label corresponding to the preprocessed signal, as well as the probability of belonging to the predicted quality label; Step 5: Based on the predicted quality label and the probability of belonging to that predicted quality label, filter all fused feature vectors to obtain multiple target fused feature vectors; Step 6: Use multiple attention residual network sub-models to process the fusion feature vector of each target to obtain the target concentration and target uncertainty parameters.

2. The method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology according to claim 1, characterized in that, Step 2 specifically includes: The preprocessed signal is subjected to time-domain feature extraction to obtain time-domain features. The preprocessed signal is subjected to fast Fourier transform to obtain the spectrum. The spectrum is then subjected to feature extraction to obtain frequency-domain features. The preprocessed signal is decomposed using the db2 wavelet basis to obtain wavelet coefficients at different scales. The energy at each scale is then calculated based on these wavelet coefficients using the following formula: ； in, For the energy at the j-th scale, For the wavelet coefficients at the j-th scale; The total wavelet energy is obtained by summing the energy at all scales. The ratio of the energy at each scale to the total wavelet energy is calculated to obtain the energy fraction. The energy fractions at the preset scales are used to form the wavelet transform features. The preprocessed signal is decomposed by empirical mode decomposition to obtain multiple intrinsic mode function components. The first three intrinsic mode function components are obtained. For each of the first three intrinsic mode function components, the decomposition characteristics of the intrinsic mode function component are calculated, including the mean, standard deviation, maximum value, minimum value, peak-to-peak value, and absolute value of the mean. The decomposition characteristics of the first three intrinsic mode function components constitute the empirical mode decomposition characteristics. The fused feature vector is obtained by fusing time-domain features, frequency-domain features, wavelet transform features, empirical mode decomposition features, and physical model-based features.

3. The method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology according to claim 2, characterized in that, The time-domain features include mean, standard deviation, root mean square (RMS), peak-to-peak value, kurtosis, skewness, zero-crossing rate, absolute maximum value, waveform factor, peak factor, and impulse factor. The absolute maximum value is the maximum amplitude in the preprocessed signal, the waveform factor is the ratio of RMS to the absolute value of the mean, the peak factor is the ratio of the absolute maximum value to RMS, and the impulse factor is the ratio of the absolute maximum value to the absolute value of the mean.

4. The method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology according to claim 2, characterized in that, The frequency domain features include: spectral centroid, spectral bandwidth, total energy, spectral roll-off point, and spectral flatness; The spectral centroid is calculated using the following formula: ； Where f i Let X(f) be the frequency. i ) represents the Fourier coefficients for the corresponding frequencies; The energy is the square of the spectral amplitude corresponding to the frequency; the spectral bandwidth is the weighted square of the difference between the frequency and the centroid of the spectrum; the total energy is the sum of the squares of the spectral amplitudes; the spectral roll-off point is the frequency in the spectrum where the energy of the frequency accumulates to 85% of the total energy from the lowest frequency to the highest frequency; and the spectral flatness is the ratio of the geometric mean to the arithmetic mean of the spectral amplitudes in the spectrum.

5. The method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology according to claim 1, characterized in that, The physical model-based features are obtained in the following ways: Experiments were conducted on the particle collision process in a solid-liquid mixture, and the following parameters were collected: rotational frequency f, concentration correction factor K, particle mass m, tangential velocity V, frictional work W, and conversion efficiency. Sensor area S and rotational speed value; Calculate the number of collisions per unit time between particles and the container wall in a solid-liquid mixture. Specifically, this is achieved through the following formula: ； Calculate the energy range of particle collisions with the container wall in a solid-liquid mixture. Specifically, this is achieved through the following formula: ； Calculate the acoustic emission energy captured by the sensor Specifically, this is achieved through the following formula: ； The mean, median, skewness, and kurtosis of the preprocessed signal are obtained to acquire the signal's statistical characteristics. Based on the sum of the squares of the rotational speed values The speed-related characteristics are calculated and expressed as follows: ； The , , The features based on the physical model consist of signal statistical characteristics and speed-related characteristics.

6. The method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology according to claim 1, characterized in that, The concentration range prediction classifier in step 3 is trained in the following way: The historical fusion feature vector is obtained as the first input sample, and the concentration range label corresponding to the historical fusion feature vector is obtained as the first output sample. The first input sample and the first output sample form the first training sample, and multiple first training samples form the first training sample set. Multiple classifiers are obtained, including SVM, KNN, Random Forest, and XGBoost. Each classifier is trained based on the first training sample set to obtain the first trained classifier. The classifier with the highest accuracy among all the first classifiers is selected as the concentration range prediction classifier.

7. The method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology according to claim 1, characterized in that, The quality classifier in step 4 is trained in the following way: A historical fusion feature vector is obtained as a second input sample, and the corresponding quality label is obtained as a second output sample. When the preprocessed signal corresponding to the historical fusion feature vector meets the abnormal conditions, the quality label is an invalid signal; when the preprocessed signal corresponding to the historical fusion feature vector does not meet the abnormal conditions, the quality label is a valid signal. The abnormal conditions include at least one of the following: signal-to-noise ratio is lower than a preset threshold, signal amplitude is lower than a preset threshold, the proportion of abnormal noise in the spectrum exceeds a preset ratio, sensor output saturation, or missing signal data. The second input sample and the second output sample constitute the second training sample, and multiple second training samples constitute the second training sample set. The second training sample set is divided into multiple sub-training sets according to the concentration range of the historical fusion feature vectors; Multiple classifiers are obtained, including SVM, KNN, Random Forest, and XGBoost. For each sub-training set, each classifier is trained using the sub-training set to obtain a trained second classifier. The classifier with the highest accuracy among all the second classifiers is selected as the quality classifier for the concentration range corresponding to the sub-training set, thus obtaining the quality classifier for each concentration range.

8. The method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology according to claim 1, characterized in that, Step 5 includes: The fusion feature vectors whose predicted quality labels match the true quality labels corresponding to the fusion feature vectors and whose probability is greater than a preset threshold are obtained and used as target fusion feature vectors. All target fusion feature vectors form the optimal feature subset.

9. The method for predicting the concentration of a solid-liquid mixture based on acoustic emission technology according to claim 1, characterized in that, Step 6 includes: By simultaneously inputting the target fused feature vector into multiple attention residual network sub-models, multiple mean predicted concentrations and multiple intra-model uncertainty parameters are obtained: Among them, the attention residual network sub-models have the same structure but different parameters; In the attention residual network sub-model, the target fusion feature vector passes through multiple sequential feature extraction modules and then through the output layer to obtain the mean predicted concentration and the uncertainty parameters within the model. The feature extraction module includes a residual block and an attention block. In the first feature extraction module, the target fused feature vector is passed sequentially through a linear layer, a batch normalization layer, a ReLU activation function, another linear layer, and another batch normalization layer to obtain transformed features. The transformed features are added to the fused feature vector, and the added features are passed through a ReLU activation function to obtain residual concatenated features. The residual concatenated features are passed sequentially through a linear layer, a ReLU activation function, a linear layer, and a Sigmoid activation function to generate attention weights. The attention weights are multiplied element-wise with the residual concatenated features to obtain weighted features. The weighted features are input to the next feature extraction module until the last feature extraction module outputs high-level features. The high-level features are input into the output layer. In the output layer, the high-level features pass through a linear layer to obtain the predicted concentration mean. At the same time, the high-level features pass through a linear layer and a Softplus activation function to obtain the standard deviation, which is used as the uncertainty parameter within the model. The arithmetic mean of the predicted concentrations output by all attention residual network sub-models is calculated to obtain the target concentration. The arithmetic mean of the intra-model uncertainty parameters output by all attention residual network sub-models is calculated to obtain the final intra-model uncertainty parameters. The standard deviation of the predicted concentrations output by all attention residual network sub-models is calculated to obtain the inter-model uncertainty parameters. The final intra-model uncertainty parameters and the inter-model uncertainty parameters constitute the target uncertainty parameters.

10. A concentration prediction system for a solid-liquid mixture based on acoustic emission technology, characterized in that, include: The signal acquisition module is used to acquire multiple acoustic emission signals during the solid-liquid mixing process; The signal preprocessing module is used to perform Butterworth bandpass filtering and normalization on each acoustic emission signal to obtain the preprocessed signal. The feature extraction and fusion module is used to obtain physical model-based features for each preprocessed signal, extract features from the preprocessed signal, and fuse them with the physical model-based features to obtain a fused feature vector. The signal filtering module is used to input the fused feature vector into the concentration range prediction classifier to obtain the initial concentration range. Based on the initial concentration range, the corresponding quality classifier is selected, and the fused feature vector is input into the quality classifier to obtain the predicted quality label corresponding to the preprocessed signal and the probability of belonging to the predicted quality label. Based on the predicted quality label and the probability of belonging to the predicted quality label, all fused feature vectors are filtered to obtain multiple target fused feature vectors. The prediction model module is used to process the fused feature vector of each target using multiple attention residual network sub-models to obtain the target concentration and target uncertainty parameters.

Images