Ball mill internal state soft measurement method based on DEM simulation and sound signal

Abstract

The application provides a ball mill internal state soft measurement method based on DEM simulation and sound signals, and the method comprises the following steps: constructing a DEM simulation model according to the physical parameters of a laboratory ball mill and performing iterative calibration; under different operating parameters, collecting sound signals by laboratory ball mill experiments, obtaining corresponding internal state data by the calibrated simulation model, and collaboratively constructing a soft measurement training data set; generating paired noise reduction module training data through continuous wavelet transform, time domain superposition and pairing processing; on this basis, training the noise reduction module and the soft measurement module, realizing noise reduction processing of the noise-containing time-frequency diagram and accurate prediction of the ball mill internal state variables. The application takes the internal state variables such as filling rate, collision energy distribution and material particle size distribution extracted by DEM simulation as training labels, and fundamentally solves the problem that the soft measurement precision is limited due to the insufficient label accuracy of the existing method.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to a soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals. Background Technology

[0002] Ball mills, as key equipment in mineral processing, account for approximately 40% to 50% of the total energy consumption of the entire mineral processing process, but their energy utilization efficiency is extremely low. Achieving refined grinding relies on real-time sensing of the mill's internal state, including key variables such as filling rate, energy distribution of collisions between the media and the wall, and particle size distribution. However, these variables cannot be directly measured during production, making soft sensing technology a primary approach to solving this problem.

[0003] Due to their non-invasiveness and ease of acquisition, acoustic signals are the mainstream input signal in the field of soft measurement for mills. Existing research has established a mapping relationship between external measurable signals and the internal state of the mill through statistical methods, machine learning, or deep learning, and has achieved certain results. However, existing methods face three fundamental limitations: First, microscopic internal state variables such as collision energy distribution cannot be obtained through mill shutdown sampling or offline calibration, resulting in a severe lack of training data with accurate internal state labels, fundamentally limiting the accuracy of soft measurement. Second, industrial ball mills operate in complex acoustic environments, with multiple noise sources such as adjacent mills, conveying equipment, and hydraulic systems superimposing each other, severely polluting the acquired acoustic signals. How to suppress industrial noise while preserving effective signal characteristics is the core challenge facing the industrial application of soft measurement systems. Third, existing deep learning-based soft measurement models rely entirely on data-driven approaches and fail to embed the known physical correspondence between the frequency components of the ball mill's acoustic signals and internal state variables into the model structure. This results in insufficient generalization ability of the model when the training samples are limited or the operating condition distribution is offset, and the prediction results lack physical rationality. Summary of the Invention

[0004] In view of the above situation, the main objective of this invention is to propose a soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals, so as to solve the above-mentioned technical problems.

[0005] This invention proposes a soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals. The method includes the following steps: Step 1: Based on the physical parameters of the laboratory ball mill, construct a three-dimensional discrete element method simulation model (DEM); iteratively calibrate the contact parameters in the three-dimensional discrete element method simulation model to obtain the calibrated discrete element method simulation model; Step 2: Conduct laboratory ball mill experiments under preset experimental operating condition parameter sets and collect experimental sound signals; run the calibrated discrete element method simulation model under preset experimental operating condition parameter sets to obtain the label set of internal state variables of the ball mill; align and integrate the sound signals with the internal state data to construct a training dataset. Step 3: Perform continuous wavelet transform on the one-dimensional time-domain sound signal in the training dataset to obtain a clean time-frequency map; introduce industrial background noise and perform time-domain superposition to obtain a noisy time-frequency map; pair the clean time-frequency map with the noisy time-frequency map to obtain paired training data. Step 4: Construct a noise reduction module based on paired training data, and construct a soft measurement module based on the clean time-frequency graph, power consumption data, and internal state variable labels; use the noise reduction module to suppress noise in the noisy time-frequency graph to obtain a noise-reduced time-frequency graph; input the noise-reduced time-frequency graph and power consumption data into the soft measurement module to obtain the prediction results of the internal state variables of the ball mill.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention uses sound signals as a soft measurement medium and extracts microscopic internal state variables such as mill filling rate, collision energy distribution, and material particle size distribution through discrete element method simulation as training labels. The mechanism data obtained by simulation is aligned and integrated with the sound signals and power consumption data collected in the experiment, realizing the organic integration of mechanism data and measured data. This fundamentally solves the problem of limited soft measurement accuracy caused by insufficient label accuracy in existing methods.

[0007] 2. This invention effectively suppresses industrial noise interference and significantly improves the prediction robustness of the system under industrial deployment conditions by superimposing on-site measured industrial background noise to train the noise reduction module and adopting a phased independent training strategy.

[0008] 3. This invention simultaneously collects sound signals and power consumption data when constructing the training dataset. The power consumption scalar data is used as an auxiliary input and fused with the sound time-frequency features in the soft measurement module for multimodal analysis. This fully utilizes the complementarity of acoustic information and energy consumption information, and effectively improves the joint prediction accuracy of filling rate, collision energy distribution and material particle size distribution. Attached Figure Description

[0009] Figure 1 This is a flowchart of the soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals proposed in this invention. Figure 2 This is a schematic diagram of the soft measurement module of the soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals proposed in this invention. Figure 3This is a schematic diagram of the noise reduction module of the soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals proposed in this invention. Detailed Implementation

[0010] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0012] Please see Figure 1 This invention proposes a soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals. The method includes the following steps: Step 1: Based on the physical parameters of the laboratory ball mill, construct a three-dimensional discrete element method simulation model; iteratively calibrate the contact parameters in the three-dimensional discrete element method simulation model to obtain the calibrated discrete element method simulation model; In step 1, a three-dimensional discrete element method simulation model is constructed based on the physical parameters of the laboratory ball mill. The contact parameters in the three-dimensional discrete element method simulation model are iteratively calibrated to obtain the calibrated discrete element method simulation model. The specific steps are as follows: Step 101: Based on the physical parameters of the laboratory ball mill, such as its geometric dimensions, cylinder rotation speed, grinding media (steel balls), filling amount, material particle size distribution, and filling amount, a three-dimensional discrete element method simulation model is established, and an initial discrete element method simulation model is obtained. The contact parameters that need to be calibrated in the three-dimensional discrete element method simulation model include: the static friction coefficient, rolling friction coefficient, and elastic recovery coefficient between steel balls; and the static friction coefficient, rolling friction coefficient, and elastic recovery coefficient between the steel balls and the cylinder wall. Step 102: Run the laboratory ball mill under at least one known operating condition, and simultaneously collect experimental measurements of macroscopically verifiable quantities: mill power consumption and material load motion morphology obtained by image analysis method using transparent end caps (characterized by the shoulder and toe position angles of the material load). Step 103: Run the initial discrete element method simulation model based on the preset experimental operating parameters and extract each simulation output value; among them, the formula for calculating the relative error of the macroscopically verifiable quantity using the experimental measurement value and simulation output value is as follows: ; in, This represents the relative error of the k-th macroscopically verifiable quantity. This represents the simulation output value of the kth macroscopically verifiable quantity. This represents the experimental measurement value of the kth macroscopically verifiable quantity; Step 104: Determine whether the relative errors of all macroscopically verifiable quantities meet the preset convergence threshold (power consumption error). Angular error of the shoulder position of the material load Toe position angle error If the preset convergence threshold is met, the current combination of contact parameters is used as the calibration result, and the calibrated discrete element method simulation model is obtained; if the preset convergence threshold is not met, the corresponding contact parameters are adjusted in turn according to the contribution of each error, with each adjustment step not exceeding 10% of the current value of the parameter, and the process returns to step 103 to recalculate the error until the convergence condition is met or the maximum number of iterations is reached, so as to obtain the calibrated discrete element method simulation model.

[0013] It should be noted that this invention uses a laboratory ball mill with a diameter of 850 mm as the research object, and conducts experiments and simulations under dry grinding conditions with a rotation speed set at 60 rpm. By adjusting the simulation contact parameters, the macroscopic verifiable quantities output by the simulation are made to match the experimental measurements, thus completing the calibration of the simulation model. The macroscopic verifiable quantities include at least one of the following: mill power consumption, mill shell vibration signal spectrum characteristics, and material load motion pattern. The material load motion pattern is obtained by using a transparent end cap combined with image analysis methods, and is characterized by the shoulder position angle and toe position angle of the material load.

[0014] Step 2: Conduct laboratory ball mill experiments under preset experimental operating condition parameter sets and collect experimental sound signals; run the calibrated discrete element method simulation model under preset experimental operating condition parameter sets to obtain the label set of internal state variables of the ball mill; align and integrate the sound signals with the internal state data to construct a training dataset. In step 2, a laboratory ball mill experiment is conducted under a preset set of experimental parameters, and experimental sound signals are collected. A calibrated discrete element method simulation model is run under the preset set of experimental parameters to obtain a label set of internal state variables of the ball mill. The sound signals and internal state data are aligned and integrated to construct a training dataset. The specific steps are as follows: Based on the ball mill filling rate, at least three discrete filling rates are preset; under each discrete filling rate, different combinations of material particle size distributions are preset to obtain the experimental working condition parameter set; the experimental working condition parameter set includes a full permutation of working conditions with at least three discrete filling rates and multiple material particle size distribution combinations, wherein the material particle size distribution is characterized by the median particle size and particle size span. Under the conditions set in the experimental operating parameter set, the experimental sound signal and power consumption data were obtained by running a laboratory ball mill. Based on the experimental operating condition parameter set, the calibrated discrete element method simulation model is used to perform simulation in order to obtain the label set of internal state variables of the ball mill. The ball mill internal state variable label set, experimental sound signals, and power consumption data were aligned and integrated to construct a training dataset. Specifically, the working condition number in the experimental working condition parameter set was used as the alignment key to integrate the experimental sound signals and power consumption data collected under the same working condition with the ball mill internal state variable labels output by the discrete element method simulation corresponding to that working condition to construct the training dataset. Each training sample consists of a sound signal segment, power consumption value, and corresponding internal state variable label triplet under the same working condition number.

[0015] It should be noted that this step involves simultaneous acquisition of experimental signals and simulation calculations under various operating conditions. In this step, a total of 131 different operating conditions were designed, with operating parameters covering four discrete filling rates (0.15, 0.22, 0.30, 0.40) and 25 different combinations of material particle size distributions. The particle size distribution is determined by the median particle size... and span Common characteristics, among which, , and These represent the particle size values ​​corresponding to a cumulative particle size distribution volume fraction of 10%, 50%, and 90%, respectively. Sound signals are collected via acoustic emission sensors or accelerometers installed on the outer surface of the mill cylinder, with the collection location at the outer surface of the cylinder corresponding to the region of strongest material collision. Simultaneously, a power meter records the mill power consumption under each operating condition. A DEM simulation with parameters consistent with this operating condition is run concurrently, extracting internal mill state variables as labels. These internal state variables include the mill filling rate, the quantile of the energy distribution of the media-wall collision (…). , , , , ) and particle size distribution quantiles ( , , , , This allows for the construction of a training dataset that corresponds one-to-one with the experimentally obtained sound signals and power consumption, and the internal state labels obtained through simulation; among which, , , , , These are the energy distribution indices corresponding to the 10%, 30%, 50%, 70%, and 90% quantiles, respectively. , , , , The granularity distributions are the 10%, 30%, 50%, 70%, and 90% quantiles, respectively.

[0016] Step 3: Perform continuous wavelet transform on the one-dimensional time-domain sound signal in the training dataset to obtain a clean time-frequency map; introduce industrial background noise and perform time-domain superposition to obtain a noisy time-frequency map; pair the clean time-frequency map and the noisy time-frequency map to obtain paired training data. In step 3, continuous wavelet transform is performed on the one-dimensional time-domain sound signal in the training dataset to obtain a clean time-frequency map; industrial background noise is introduced and time-domain superposition is performed to obtain a noisy time-frequency map; the clean time-frequency map and the noisy time-frequency map are paired to obtain paired training data. The specific steps are as follows: Step 301: Using the Morlet wavelet as the target wave function, perform continuous wavelet transform on the one-dimensional time-domain sound signal of each sample in the training dataset to obtain wavelet coefficients; after traversing different scale parameters and translation parameters, map the modulus of the wavelet coefficients to a two-dimensional image matrix to obtain a clean time-frequency diagram. Step 302: Obtain the industrial background noise signal; perform time-domain superposition of the industrial background noise signal with the one-dimensional time-domain sound signal of each sample in the training dataset to obtain the superimposed noisy sound signal; Step 303: The superimposed noisy audio signal is used as input, and Morlet wavelet is used as the target wave function. After traversing different scale parameters and translation parameters, the modulus of the wavelet coefficients is mapped to a two-dimensional image matrix to obtain the noisy time-frequency map. Step 304: Pair the clean time-frequency map and the noisy time-frequency map obtained from the one-dimensional time-domain sound signal of the same sample, use the clean time-frequency map as the training target and the noisy time-frequency map as the training input to obtain paired training data.

[0017] Specifically, in the process of using the Morlet wavelet as the target wavelet function to perform continuous wavelet transform on the one-dimensional time-domain sound signal of each sample in the training dataset to obtain the wavelet coefficients, the following relationship exists: ; in, Indicated in scale parameter With translation parameters The wavelet coefficients are used to reflect the energy intensity of the signal at that time-frequency position; This represents the one-dimensional time-domain sound signal to be analyzed at time t. Represents the complex conjugate of the mother wavelet function. Represents an integral infinitesimal element; The scaling parameter is used to control the scaling of the wavelet and is inversely proportional to the frequency. This represents the translation parameter, used to control the position of the wavelet along the time axis.

[0018] Step 4: Construct a noise reduction module based on paired training data, and construct a soft measurement module based on the clean time-frequency graph, power consumption data, and internal state variable labels; use the noise reduction module to suppress noise in the noisy time-frequency graph to obtain a noise-reduced time-frequency graph; input the noise-reduced time-frequency graph and power consumption data into the soft measurement module to obtain the prediction results of the internal state variables of the ball mill. Please see Figure 2 and Figure 3 In step 4, a denoising module is constructed based on paired training data, and a soft measurement module is constructed based on the clean time-frequency graph, power consumption data, and internal state variable labels. The denoising module is used to suppress noise in the noisy time-frequency graph to obtain a denoised time-frequency graph. The denoised time-frequency graph and power consumption data are input into the soft measurement module to obtain the prediction results of the internal state variables of the ball mill. The specific steps are as follows: A noise reduction module is constructed based on the encoder-decoder structure. The noise reduction module consists of a three-layer downsampling encoder, a bottleneck layer, and a three-layer upsampling decoder cascaded together, with a skip connection between the encoder and the decoder. The noisy time-frequency image is compressed layer by layer by the encoder, and then the global context is modeled in the channel dimension by the multi-depth convolutional head transpose attention mechanism in the bottleneck layer. The decoder then recovers the noisy time-frequency image layer by layer and outputs the predicted noise component. The noisy time-frequency image is obtained by subtracting the predicted noise component from the noisy time-frequency image. Using noisy and clean time-frequency maps from paired training data, the network parameters of the denoising module are iteratively trained using mini-batch gradient descent until the loss function converges, thus obtaining a preliminarily trained denoising module. After evaluating the performance of the initially trained noise reduction module, all network parameters of the initially trained noise reduction module are frozen to obtain a noise reduction module with fixed parameters. A soft measurement module is constructed based on Vision Transformer. The pure time-frequency graph is divided into image blocks of equal size. Each image block is linearly projected and superimposed with a physical mechanism-enhanced position code to obtain the superimposed sequence. The soft measurement module is based on Vision Transformer and makes two improvements for the soft measurement task of the ball mill: First, the standard position code is replaced with a physical mechanism-enhanced position code to embed the physical correspondence between the frequency components of the sound signal and the internal state variables. Second, after feature extraction, multimodal fusion technology is used to merge the power consumption features with the time-frequency graph features to improve the accuracy of soft measurement. The superimposed sequence is input into the Transformer encoder, and global feature interaction is performed through a multi-head self-attention mechanism to obtain the global features of the time-frequency map. Specifically, each image block interacts with image blocks at all locations in the world through the multi-head self-attention mechanism to capture the long-range dependencies between different frequency locations in the time-frequency map. The multi-head parallel output is stitched and linearly projected to obtain the global features of the time-frequency map. The power consumption data in the training dataset is normalized and then mapped to a power embedding vector through a linear layer. The global features of the time-frequency map and the power embedding vector are concatenated along the feature dimension and then mapped through a multilayer perceptron fusion network consisting of multiple fully connected layers and activation functions to obtain the soft measurement value. A target function is constructed using soft measurement values. Using the clean time-frequency plot, power consumption data, and ball mill internal state variable label set in the training data, all trainable parameters in the soft measurement module are optimized through the target function to obtain a trained soft measurement module. The noisy time-frequency graph is input into a noise reduction module with fixed parameters. The noise components are predicted by forward propagation, and the noise-reduced time-frequency graph is obtained. The noise-reduced time-frequency graph and the corresponding power consumption data are input into the trained soft measurement module to obtain preliminary measurement results. The coefficient of determination and mean absolute error are used as performance indicators to evaluate the preliminary measurement results. The preliminary measurement results that pass the evaluation are used as the prediction results of the internal state variables of the ball mill.

[0019] Specifically, in the process of constructing the noise reduction module based on the encoder-decoder structure, the training objective of the noise reduction module is to make the noise components predicted by the encoder-decoder network approximate the actual noise components. The relationship of the training objective is as follows: ; in, Represents the true noise component. This represents a noisy time-frequency graph. This represents a clean time-frequency diagram. This represents the time-frequency graph after noise reduction. This represents the noise component predicted by the noise reduction module. Using noisy and clean time-frequency maps from paired training data, the network parameters of the denoising module are iteratively trained using mini-batch gradient descent until the loss function converges, thus obtaining a preliminarily trained denoising module. The expression for the loss function is: ; in, This represents the value of the loss function; It represents the square of the Euclidean norm, which is the sum of the squares of each element of the difference between the predicted noise component and the actual noise component, and is used to measure the overall deviation between the two. A soft measurement module is built based on Vision Transformer; the clean time-frequency image is divided into image blocks of equal size; each image block is linearly projected and then superimposed with the physical mechanism-enhanced positional encoding to obtain the superimposed sequence, where the expression for the physical mechanism-enhanced positional encoding is: ; in, Indicates the first Physical mechanism-enhanced location coding corresponding to each image patch Indicates the first Standard location encoding of image blocks Indicates the first The frequency center value corresponding to each image patch; The frequency physical weight vector is initialized based on the following physical mechanism: the low-frequency component of the ball mill sound signal carries filling rate information, while the high-frequency component reflects the collision energy information of the medium. Therefore, the bias vector elements corresponding to the low-frequency region image patch are initialized to positive values, so that the location encoding of this region is positively enhanced in the dimension of filling rate-related features, thereby strengthening the association with filling rate prediction; the bias vector elements corresponding to the high-frequency region image patch are initialized to negative values, so that the location encoding of this region is negatively suppressed in the dimension of collision energy-related features, thereby strengthening the association with collision energy distribution prediction; the above initial values ​​are adaptively adjusted with data during network training. This represents the learnable scaling parameters jointly trained with the soft measurement module, when the physical mechanism guidance is effective. The trend tends to increase when the physical mechanism guides the data but is inconsistent. Approaching zero, the position code degenerates into a standard form, thereby enabling adaptive adjustment of the physical mechanism-guided intensity; The superimposed sequence is input into the Transformer encoder, and global feature interaction is performed through a multi-head self-attention mechanism to obtain the global features of the time-frequency map. Specifically, each image block interacts with image blocks at all locations in the world through the multi-head self-attention mechanism to capture the long-range dependencies between different frequency locations in the time-frequency map. The multi-head parallel output is stitched and linearly projected to obtain the global features of the time-frequency map. Construct an objective function using soft measurement values, where the expression for the objective function is: ; in, Represents the objective function value. Indicates the total number of image patches. Indicates the first The true values ​​of the internal state variables of the ball mill corresponding to each image patch; The model is represented by the first Each image block is taken as input, and the soft measurement value output is the corresponding internal state variable of the ball mill for each input image block; Indicates the first The squared error between the true value and the soft measurement estimate corresponding to each image patch; The denoised time-frequency graph and corresponding power consumption data are input into the trained soft measurement module to obtain preliminary measurement results. The coefficient of determination and mean absolute error are used as performance indicators to evaluate the preliminary measurement results. After passing the evaluation, the prediction results of the internal state variables of the ball mill are obtained. The formulas for calculating the coefficient of determination and mean absolute error are as follows: ; in; The coefficient of determination is denoted by 1, and the closer it is to 1, the better the model fit. The mean of the true values; It represents the mean absolute error, used to reflect the average deviation between the soft measurement value and the true value.

[0020] It should be noted that in this step, the soft measurement module uses the Vision Transformer network to divide the time-frequency map into several image blocks of equal size. Each image block is linearly projected and superimposed with the physical mechanism-enhanced position code, which is then input into the Transformer encoder. Global feature interaction is achieved through a multi-head self-attention mechanism.

[0021] The performance of the soft measurement module before and after noise reduction was compared, and the specific results are shown in Table 1: Table 1: Comparison of Soft Measurement Performance Before and After Noise Reduction

[0022] Table 1 shows that the soft measurement accuracy of all variables improved after noise reduction, with the improvement varying significantly depending on the variable type. The fill rate was least affected by noise; R² increased from 0.95 to 0.96 before and after noise reduction, while the MSE remained essentially unchanged. (Collision energy distribution) The MSE was uniformly increased from 0.82~0.83 to 0.90~0.91, resulting in a decrease of approximately 40%. The most significant improvement was in the particle size distribution. The MSE decreased by about 80%, from 0.93~0.94 to 0.98~0.99, indicating that the granularity distribution information is most sensitive to industrial noise interference. At the same time, it shows that the noise reduction module of the present invention significantly improves the prediction robustness of the system by effectively restoring the detailed structure of the time-frequency map.

[0023] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0024] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0025] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals, characterized in that, The method includes the following steps: Step 1: Based on the physical parameters of the laboratory ball mill, construct a three-dimensional discrete element method simulation model; iteratively calibrate the contact parameters in the three-dimensional discrete element method simulation model to obtain the calibrated discrete element method simulation model; Step 2: Conduct laboratory ball mill experiments under preset experimental operating condition parameter sets, and collect experimental sound signals and power consumption data; run the calibrated discrete element method simulation model under preset experimental operating condition parameter sets to obtain the ball mill internal state variable label set; align and integrate the ball mill internal state variable label set, experimental sound signals, and power consumption data to construct a training dataset. Step 3: Perform continuous wavelet transform on the one-dimensional time-domain sound signals in the training dataset to obtain a clean time-frequency diagram; By introducing industrial background noise and performing time-domain superposition, a noisy time-frequency map is obtained; the clean time-frequency map and the noisy time-frequency map are paired to obtain paired training data. Step 4: Construct a noise reduction module based on paired training data, and construct a soft measurement module based on clean time-frequency plots, power consumption data, and internal state variable labels; The noise reduction module is used to suppress noise in the noisy time-frequency graph to obtain a denoised time-frequency graph. The denoised time-frequency graph and power consumption data are then input into the soft measurement module to obtain the prediction results of the internal state variables of the ball mill. The specific steps are as follows: A noise reduction module is constructed based on the encoder-decoder structure; A soft measurement module is constructed based on Vision Transformer. Based on the soft measurement module, a sliding window with a fixed step size is used to uniformly divide the pure time-frequency image into image blocks of equal size. Each image block is then linearly projected and superimposed with the position code of physical mechanism enhancement to obtain the superimposed sequence. The superimposed sequence is input into the Transformer encoder, and global feature interaction is performed through a multi-head self-attention mechanism to obtain global features of the time-frequency map; The power consumption data in the training dataset is normalized and then mapped to a power embedding vector through a linear layer. The global features of the time-frequency map and the power embedding vector are concatenated along the feature dimension and then mapped through a multilayer perceptron fusion network consisting of multiple fully connected layers and activation functions to obtain the soft measurement value. Construct an objective function using soft measurement values; Using the clean time-frequency plot, power consumption data, and ball mill internal state variable label set in the training data, all trainable parameters in the soft measurement module are optimized through the objective function to obtain a trained soft measurement module. The noisy time-frequency graph is input into a noise reduction module with fixed parameters. The noise components are predicted by forward propagation, and the noise-reduced time-frequency graph is obtained. The denoised time-frequency graph and the corresponding power consumption data are input into the trained soft measurement module to obtain preliminary measurement results; the preliminary measurement results after evaluation are used as the prediction results of the internal state variables of the ball mill. The location coding expression for physical mechanism enhancement is: ； Indicates the first Physical mechanism-enhanced location coding corresponding to each image patch Indicates the first Standard location encoding of image blocks Indicates the first The frequency center value corresponding to each image patch; This represents the frequency physical weight vector, which is initialized based on physical mechanisms. It is used to initialize the bias vector elements corresponding to low-frequency region image blocks to positive values ​​and the bias vector elements corresponding to high-frequency region image blocks to negative values. This represents the learnable scaling parameters that are trained together with the soft measurement module.

2. The soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals according to claim 1, characterized in that, In step 1, a three-dimensional discrete element method simulation model is constructed based on the physical parameters of the laboratory ball mill; the contact parameters in the three-dimensional discrete element method simulation model are iteratively calibrated to obtain the calibrated discrete element method simulation model. The specific steps are as follows: Step 101: Based on the physical parameters of the laboratory ball mill, establish a three-dimensional discrete element method simulation model and obtain an initial discrete element method simulation model; among which, the contact parameters that need to be calibrated in the three-dimensional discrete element method simulation model include: the static friction coefficient between steel balls, the static friction coefficient between steel balls and the cylinder wall, the rolling friction coefficient, and the elastic recovery coefficient. Step 102: Run the laboratory ball mill under at least one known operating condition, while collecting experimental measurements of macroscopically verifiable quantities. Step 103: Run the initial discrete element method simulation model based on the preset experimental operating parameters and extract each simulation output value; among them, the formula for calculating the relative error of the macroscopically verifiable quantity using the experimental measurement value and simulation output value is as follows: ； in, This represents the relative error of the k-th macroscopically verifiable quantity. This represents the simulation output value of the kth macroscopically verifiable quantity. This represents the experimental measurement value of the kth macroscopically verifiable quantity; Step 104: Determine whether the relative errors of all macroscopic verifiable quantities meet the preset convergence threshold. If the preset convergence threshold is met, the current combination of contact parameters is used as the calibration result, and the calibrated discrete element method simulation model is obtained. If the preset convergence threshold is not met, the corresponding contact parameters are adjusted in sequence according to the magnitude of their contribution to each error, so as to obtain the calibrated discrete element method simulation model.

3. The soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals according to claim 2, characterized in that, In step 2, a laboratory ball mill experiment is conducted under a preset set of experimental parameters, and experimental sound signals and power consumption data are collected. A calibrated discrete element method simulation model is run under the preset set of experimental parameters to obtain a label set of internal state variables of the ball mill. The label set of internal state variables of the ball mill, the experimental sound signals, and the power consumption data are aligned and integrated to construct a training dataset. The specific steps are as follows: Based on the ball mill filling rate, at least three discrete filling rates are preset; under each discrete filling rate, different combinations of material particle size distribution are preset to obtain the experimental operating condition parameter set. Under the conditions set in the experimental operating parameter set, the experimental sound signal and power consumption data were obtained by running a laboratory ball mill. Based on the experimental operating condition parameter set, the calibrated discrete element method simulation model is used to perform simulation in order to obtain the label set of internal state variables of the ball mill. The ball mill's internal state variable label set, experimental sound signals, and power consumption data were aligned and integrated to construct a training dataset.

4. The soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals according to claim 3, characterized in that, In step 3, a continuous wavelet transform is performed on the one-dimensional time-domain sound signal in the training dataset to obtain a clean time-frequency map; industrial background noise is introduced and time-domain superposition is performed to obtain a noisy time-frequency map; the clean time-frequency map and the noisy time-frequency map are paired to obtain paired training data. The specific steps are as follows: Step 301: Using the Morlet wavelet as the target wave function, perform continuous wavelet transform on the one-dimensional time-domain sound signal of each sample in the training dataset to obtain wavelet coefficients; After traversing different scale and translation parameters, the modulus of the wavelet coefficients is mapped to a two-dimensional image matrix to obtain a clean time-frequency diagram. Step 302: Obtain the industrial background noise signal; perform time-domain superposition of the industrial background noise signal with the one-dimensional time-domain sound signal of each sample in the training dataset to obtain the superimposed noisy sound signal; Step 303: The superimposed noisy audio signal is used as input, and Morlet wavelet is used as the target wave function. After traversing different scale parameters and translation parameters, the modulus of the wavelet coefficients is mapped to a two-dimensional image matrix to obtain a noisy time-frequency map. Step 304: Pair the clean time-frequency map and the noisy time-frequency map obtained from the one-dimensional time-domain sound signal of the same sample, use the clean time-frequency map as the training target and the noisy time-frequency map as the training input to obtain paired training data.

5. The soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals according to claim 4, characterized in that, When using the Morlet wavelet as the target wavelet function to perform continuous wavelet transform on the one-dimensional time-domain sound signal of each sample in the training dataset to obtain the wavelet coefficients, the following relationship exists: ； in, Indicated in scale parameter With translation parameters The wavelet coefficients below, This represents the one-dimensional time-domain sound signal to be analyzed at time t. Represents the complex conjugate of the mother wavelet function. This represents an integral infinitesimal element.

6. The soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals according to claim 5, characterized in that, In step 4, a noise reduction module is constructed based on paired training data, and a soft measurement module is constructed based on clean time-frequency plots, power consumption data, and internal state variable labels. A noise reduction module is used to suppress noise in the noisy time-frequency graph to obtain a denoised time-frequency graph. The denoised time-frequency graph and power consumption data are then input into a soft measurement module to obtain the prediction results of the internal state variables of the ball mill. The noise reduction module consists of a cascaded three-layer downsampling encoder, a bottleneck layer, and a three-layer upsampling decoder, with a skip connection between the encoder and the decoder. After the noisy time-frequency graph is compressed layer by layer by the encoder, the global context is modeled in the channel dimension by the multi-depth convolutional head transpose attention mechanism in the bottleneck layer, and then restored layer by layer by the decoder to output the predicted noise component. The denoised time-frequency graph is obtained by subtracting the predicted noise component from the noisy time-frequency graph. The network parameters of the denoising module are iteratively trained using the mini-batch gradient descent method until the loss function converges, so as to obtain the denoising module that has been initially trained. After evaluating the performance of the initially trained noise reduction module, all network parameters of the initially trained noise reduction module are frozen to obtain a noise reduction module with fixed parameters.

7. The soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals according to claim 6, characterized in that, In the process of constructing a noise reduction module based on the encoder-decoder structure, the training objective of the noise reduction module is to make the noise components predicted by the encoder-decoder network approximate the actual noise components. The relationship of the training objective is as follows: ； in, Represents the true noise component. This represents a noisy time-frequency graph. This represents a clean time-frequency diagram. This represents the time-frequency graph after noise reduction. This represents the noise component predicted by the noise reduction module.

8. The soft measurement method for the internal state of a ball mill based on DEM simulation and sound signals according to claim 7, characterized in that, Using noisy and clean time-frequency maps from paired training data, the network parameters of the denoising module are iteratively trained using mini-batch gradient descent until the loss function converges, thus obtaining a preliminarily trained denoising module. The expression for the loss function is: ； in, This represents the value of the loss function. Represents the square of the Euclidean norm; Constructing an objective function using soft measurement values: ； in, Represents the objective function value. Indicates the total number of image patches. Indicates the first The true values ​​of the internal state variables of the ball mill corresponding to each image patch; The model is represented by the first When an image block is used as input, the soft measurement value output is the internal state variable of the ball mill corresponding to the input image block; Indicates the first The squared error between the true value and the soft measurement estimate corresponding to each image patch.

Images