Continuous ball mill power saving monitoring system

Abstract

The present invention relates to a continuous ball mill power saving monitoring system, belonging to the technical field of energy consumption monitoring. It includes a data acquisition module, a data processing module, a parameter prediction module and an energy consumption calculation module. The data processing module is used to extract features from the collected vibration signals and grinding sound signals and construct feature vectors; the parameter prediction module is used to construct a neural network model to predict the grinding concentration by inputting the vibration signals and grinding sound signals collected in real time; the energy consumption calculation module is used to characterize the energy consumption changes by using a polynomial regression model through the predicted grinding concentration, set an alarm threshold, and monitor the energy consumption changes online. The present invention establishes a grinding concentration prediction model through the collected vibration signals and grinding sound signals of the cylinder to obtain predicted values, and then inputs the predicted values ​​into the established polynomial regression model to convert and obtain the changing trend of energy consumption, thereby realizing real-time monitoring of the energy consumption of the ball mill, improving the automation level of the system, and making energy consumption monitoring more efficient.

Description

Technical Field

[0001] The invention belongs to the technical field of energy consumption monitoring, and in particular relates to a continuous ball mill power saving monitoring system. Background Art

[0002] Compared with intermittent ball mills, the biggest advantage of continuous ball mills is that the whole set of equipment adopts PLC control, which basically realizes centralized control and unmanned operation, has the best grinding effect, greatly reduces power consumption, and has obvious energy-saving effects.

[0003] However, in actual production lines, the power consumption of ball mills still accounts for a considerable proportion of the cost, so energy saving of ball mills is of great significance. At present, traditional ball mills are usually not equipped with advanced monitoring and control systems, making it difficult to monitor energy consumption in real time, making it difficult to discover and solve the problem of excessive energy consumption, thereby affecting production costs and equipment operation efficiency. At the same time, since energy saving of ball mills is a comprehensive process involving many factors, it is often not possible to achieve ideal results by only focusing on one aspect, and it is impossible to accurately grasp the overall energy consumption of the equipment during operation, which easily leads to excessive energy consumption. Summary of the invention

[0004] In order to solve the above-mentioned problems existing in the prior art, the present invention provides a continuous ball mill power saving monitoring system, which establishes a grinding concentration prediction model by collecting the vibration signal and grinding sound signal of the cylinder to obtain the predicted value, and then inputs the predicted value into the established polynomial regression model to convert it into the changing trend of energy consumption, thereby realizing real-time monitoring of the energy consumption of the ball mill, improving the degree of automation of the system, and making energy consumption monitoring more efficient.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] The present disclosure provides a continuous ball mill power saving monitoring system, comprising a data acquisition module, a data processing module, a parameter prediction module and an energy consumption calculation module, wherein the data acquisition module, the data processing module, the parameter prediction module and the energy consumption calculation module are sequentially connected in communication;

[0007] The data processing module is used to extract features from the collected vibration signals and grinding sound signals and construct feature vectors; the parameter prediction module is used to construct a neural network model to predict the grinding concentration by inputting the real-time collected vibration signals and grinding sound signals; the energy consumption calculation module is used to characterize the energy consumption changes by using a polynomial regression model through the predicted grinding concentration, set an alarm threshold, and monitor the energy consumption changes online;

[0008] The parameter prediction module predicts the grinding concentration and includes the following steps:

[0009] S21. Establishing a data set: selecting test materials, ball stones and medium filling rate, taking grinding concentration as a variable, giving different grinding concentrations to represent the load change of the ball mill, collecting vibration signals and grinding sound signals under different grinding concentrations, and establishing a data set;

[0010] S22, constructing a prediction model: using a BP neural network model to predict grinding concentration, setting an input layer according to the constructed input feature vector, taking grinding concentration as output, and establishing a grinding concentration prediction model;

[0011] S23, parameter prediction: the vibration signal and grinding sound signal obtained by the data acquisition module are processed by the data processing module and then input into the trained prediction model to obtain the grinding concentration.

[0012] As a preferred technical solution of the present invention, the data processing module constructs the feature vector including the following steps:

[0013] S11, adaptive VMD algorithm decomposition: the collected cylinder vibration signal and grinding sound signal are decomposed by adaptive VMD algorithm to obtain their respective IMF components;

[0014] S12, selecting effective IMF components: further screening the IMF components obtained after adaptive VMD decomposition by using the correlation coefficient method to obtain effective IMF components;

[0015] S13, Hilbert transform and marginal spectrum sample entropy calculation: The selected effective IMF components are subjected to Hilbert transform to obtain the instantaneous frequency and amplitude envelope of each IMF; then the marginal spectrum sample entropy of each IMF component is calculated;

[0016] S14. Construct feature vector: combine the sample entropy values ​​of all valid IMF components into a feature vector as the input of the prediction model.

[0017] As a preferred technical solution of the present invention, the adaptive VMD algorithm decomposition comprises the following steps:

[0018] Initialization parameters: Set the initial mode number K, penalty parameter α and convergence tolerance of vibration signal and grinding sound signal decomposition respectively ;

[0019] Iterative update: decompose the signal into several modes with limited frequency bands, so that the center frequency and bandwidth of each mode continue to converge to a stable state;

[0020] Decomposition results: After the iteration is completed, the IMF components of the vibration signal and the grinding sound signal are obtained. Each component reflects the dynamic characteristics of different frequency components in the signal.

[0021] The selecting of effective IMF components comprises the following steps:

[0022] Calculate the correlation coefficient: Calculate the correlation coefficient between each IMF component and the original signal to evaluate its representativeness to the original signal;

[0023] Set threshold: Set a correlation coefficient threshold according to the actual situation;

[0024] Screening IMF components: Select IMF components whose correlation coefficients are greater than the set threshold as valid components.

[0025] As a preferred technical solution of the present invention, the Hilbert transform and marginal spectrum sample entropy calculation include the following steps:

[0026] Hilbert transform: Perform Hilbert transform on each effective IMF component to obtain the instantaneous amplitude and instantaneous frequency of each component;

[0027] Marginal spectrum calculation: Combine the instantaneous amplitude and instantaneous frequency to obtain the marginal spectrum of each IMF component;

[0028] Sample entropy calculation: Calculate the sample entropy for the marginal spectrum of each IMF component;

[0029] The constructing of the feature vector comprises the steps of:

[0030] Construction of feature vector: The sample entropy values ​​calculated from the marginal spectrum of each IMF component are arranged in sequence according to the IMF sequence number to form a multi-dimensional feature vector. The dimension of the vector depends on the number of valid IMFs screened out.

[0031] Normalization: Normalize the feature vectors to make the value range of each feature consistent.

[0032] As a preferred technical solution of the present invention, the construction of the prediction model includes the following steps:

[0033] S221, model parameter setting: set input layer nodes according to the dimension of the feature vector, the number of input layer nodes is N*M, where N is the number of IMF components and M is the number of features of each component;

[0034] The prediction target is grinding concentration, and the number of nodes in the output layer is set to 1;

[0035] The empirical expression is used to calculate the number of neurons in the hidden layer. The formula is:

[0036] ;

[0037] Where: q and p is the total number of neurons in the hidden layer and the output layer.

[0038] As a preferred technical solution of the present invention, the construction of the prediction model further includes the following steps:

[0039] S222, model optimization: using a genetic algorithm to optimize the weights and thresholds of the neural network, including the following steps:

[0040] S2221, Initialization: Use the feedforwardnet function to build a 3-layer neural network, determine the weights and thresholds, and express the weights and thresholds in the form of a matrix:

[0041] Construct a weight matrix between the input layer and the hidden layer , represents the connection weight between two layers;

[0042] The threshold matrix of the hidden layer is denoted as , which includes q Row 1 column data;

[0043] Constructing weight matrix between hidden layer and output layer , represents the connection weight between two layers;

[0044] The threshold matrix of the output layer is , which includes a data;

[0045] Where n is the number of input layer nodes;

[0046] according to , , , The data in the matrix are encoded in the order of arrangement to form a gene sequence;

[0047] S2222, Adaptability evaluation: Calculate the fitness values ​​of all individuals according to the fitness function, and determine the population to continue the next round of iteration based on the fitness values. The higher the fitness value, the greater the probability of being selected.

[0048] S2223, performing crossover and mutation operations on all gene sequences according to the designed probabilities, and determining the probability intervals of the crossover and mutation operators to be [0.90, 0.99] and [0.05, 0.15] respectively according to the calculation results;

[0049] S2224, input the training data into the network, calculate the error according to the preset error function, update the weights through the forward and back propagation algorithms, and continuously optimize the parameters and structure of the network;

[0050] Use the iterative function to complete the network training and set the learning accuracy. When the network prediction error is less than the learning accuracy, the network training ends and the network weights and thresholds remain unchanged.

[0051] As a preferred technical solution of the present invention, the fitness function is calculated as follows:

[0052] ;

[0053] In the formula, is the fitness function, x For the gene sequence, K is the grinding concentration amplitude, m is the total number of training data, For sample i The actual grinding concentration, For sample i Enter the predicted values ​​obtained by the model.

[0054] As a preferred technical solution of the present invention, the energy consumption calculation module characterizes the energy consumption change and includes the following steps:

[0055] S31. Data collection and preparation: Collect historical data, including grinding concentration and corresponding energy consumption values. The data covers different working conditions and concentration ranges. Preprocess the data, including removing outliers and filling missing values.

[0056] S32, establish a polynomial regression model: select the model order, and convert energy consumption E Grinding concentration C The relationship is expressed as a polynomial function:

[0057] ;

[0058] In the formula, , ,… are the polynomial regression coefficients, m is the polynomial order;

[0059] S33. Model fitting: Use the PolynomialFeatures function to fit the polynomial model and obtain the polynomial coefficients.

[0060] As a preferred technical solution of the present invention, the characterization of energy consumption changes further includes the steps of:

[0061] S34, Model evaluation: Evaluate the model's fitting effect, including residual analysis, R² value and mean square error. When the model is overfitted, regularization techniques are used to constrain high-order coefficients.

[0062] S35. Constructing the energy consumption change trend: inputting the predicted grinding concentration into the polynomial regression model to obtain the energy consumption characterization value, and characterizing the energy consumption change trend by establishing the energy consumption change curve.

[0063] As a preferred technical solution of the present invention, the data acquisition module is used to collect vibration signals and grinding sound signals in real time, and transmit the collected signals to the data processing module through a data acquisition card, wherein the vibration signal is collected by an acceleration sensor, and the grinding sound signal is collected by a microphone.

[0064] The beneficial effects of the present invention are:

[0065] The present invention firstly takes the load state of the main controllable factor affecting the energy consumption of the continuous ball mill as the characterization basis of the energy consumption change, monitors the grinding concentration, and then uses the collected vibration signal and grinding sound signal of the cylinder as the basis for judging the grinding concentration. A grinding concentration prediction model is established to obtain the predicted value, and then the predicted value is input into the established polynomial regression model to convert it into the energy consumption change trend, thereby realizing real-time monitoring of the energy consumption of the ball mill, improving the automation degree of the system, and making energy consumption monitoring more efficient.

[0066] In the process of processing collected data, the present invention adopts adaptive VMD and Hilbert marginal spectrum sample entropy theory to construct feature vectors, specifically solves the problems of signal nonlinearity, non-stationarity and multi-coupling characteristics, reduces the influence of noise on signal analysis, and improves the accuracy of grinding concentration prediction; at the same time, in the training of the constructed BP neural network model, the weights and thresholds of the neural network are optimized by adopting a genetic algorithm, thereby ensuring the training effect of the model, avoiding overfitting of the weights and thresholds, improving the convergence speed, and improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0068] Figure 1 A schematic structural diagram of a continuous ball mill power saving monitoring system provided by an embodiment of the present invention;

[0069] Figure 2 A schematic diagram of a flow chart of steps executed by a data processing module according to an embodiment of the present invention;

[0070] Figure 3 A schematic diagram of a flow chart of execution steps of a parameter prediction module provided in an embodiment of the present invention;

[0071] Figure 4 A schematic flow chart of the steps executed by the energy consumption calculation module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0073] During the operation of the ball mill, its current is basically a curve from large to small, but the difference is only about 5%, which is basically a machine with constant torque properties. In order to achieve a better ball milling effect, the speed of the ball mill should be a certain value. Within this certain value range, the input power of the ball mill is not wasted, so the energy saving effect achieved by adjusting the speed is generally good. Therefore, in this embodiment, by monitoring the operating state of the continuous ball mill, the optimization target is excavated to ensure that the ball mill operates under rated conditions, that is, the optimal state (load) operation is achieved to achieve energy saving control.

[0074] Grinding concentration usually refers to the solid-liquid ratio of the mixture of ore and water in the ball mill, that is, the percentage of the mass of solid materials in the slurry to the total mass, which directly affects the grinding process of the ball mill:

[0075] When the grinding concentration is too low, the fluidity of the slurry is greater, the collision and friction between the material particles are weakened, the grinding efficiency is reduced, and it is easy to cause the slurry to overflow, increasing unnecessary energy consumption; when the grinding concentration is too high, the viscosity of the slurry increases, hindering the movement of the steel balls and the transmission of the impact force, and the grinding efficiency will also be reduced. At the same time, too high a concentration may cause the ball mill to overload, increase the mechanical wear and energy consumption of the equipment; at the appropriate grinding concentration, the collision and grinding of the material and the medium inside the ball mill are most effective, the grinding efficiency is the highest, and the energy consumption is the lowest.

[0076] Therefore, grinding concentration directly affects the energy transfer efficiency of the ball mill. Too high or too low concentration will lead to a decrease in energy utilization and increase the energy consumption per unit product. At the appropriate grinding concentration, the material can be fully ground and the energy consumption is also the most reasonable. Therefore, grinding concentration, as a load parameter, can directly reflect the energy consumption state of the ball mill.

[0077] Specifically, the continuous ball mill power saving monitoring system of the present invention is described in detail through the following contents:

[0078] This embodiment provides a continuous ball mill power saving monitoring system. Figure 1 As shown, it includes a data acquisition module, a data processing module, a parameter prediction module and an energy consumption calculation module, and the data acquisition module, the data processing module, the parameter prediction module and the energy consumption calculation module are communicatively connected in sequence.

[0079] The data acquisition module is used to collect vibration signals and grinding sound signals in real time, and transmit the collected signals to the data processing module through the data acquisition card, wherein the vibration signal is collected by the acceleration sensor, and the grinding sound signal is collected by the microphone.

[0080] It should be noted that, in this embodiment, since the load state of the ball mill is closely related to the movement of the ore and the steel balls, the vibration signal and sound signal generated by the steel balls hitting the cylinder when the mill rotates contain the load state information. Therefore, in this embodiment, the load state of the ball mill is indirectly judged by collecting the vibration signal and sound signal of the cylinder. It will be more accurate to use a combination of vibration signal and sound signal for judgment.

[0081] The data processing module is used to extract features from the collected vibration signals and grinding sound signals and construct feature vectors, such as Figure 2 As shown, the following steps are included:

[0082] S11, adaptive VMD algorithm decomposition: the collected cylinder vibration signal and grinding sound signal are decomposed by the adaptive VMD algorithm to obtain their respective IMF components, including the following steps:

[0083] Initialization parameters: Set the initial mode number K, penalty parameter α and convergence tolerance of vibration signal and grinding sound signal decomposition respectively ;

[0084] Iterative update: decompose the signal into several modes with limited frequency bands, so that the center frequency and bandwidth of each mode continue to converge to a stable state;

[0085] Decomposition results: After the iteration is completed, the IMF components of the vibration signal and the grinding sound signal are obtained. Each component reflects the dynamic characteristics of different frequency components in the signal.

[0086] The adaptive VMD algorithm is a non-recursive decomposition method that can decompose complex signals into several intrinsic mode functions (IMFs). Each IMF represents an intrinsic vibration mode in the signal. The main goal of VMD is to minimize the bandwidth of each IMF while maintaining the purity and non-overlapping characteristics of these modal signals.

[0087] It should be noted that due to the characteristic differences between the cylinder vibration signal and the sound signal, the VMD parameters need to be set separately when performing VMD decomposition, such as:

[0088] VMD parameter settings for vibration signals:

[0089] Number of modes K: Vibration signals usually contain low-frequency equipment vibration characteristics and some mechanical resonance frequencies. The number of modes K can be relatively small (such as 3 to 5) to capture the main vibration characteristics. Penalty parameter α: Setting a smaller α value can enhance the sensitivity of the decomposition and capture significant changes in the vibration signal (such as impact). Convergence tolerance : Since the characteristic changes of vibration signals are relatively obvious, the convergence tolerance can be set larger to ensure that the algorithm converges within a reasonable time.

[0090] VMD parameter settings for sound signals:

[0091] Number of modes K: Sound signals have complex spectral characteristics, so a larger K value is needed to decompose multiple frequency components (such as 6 to 10) to capture high-frequency noise and various sound components. Penalty parameter α: Sound signals often have smoother amplitude changes and more noise, so a larger α value is needed to enhance the robustness of VMD and avoid over-decomposition and noise amplification. Convergence tolerance :The frequency components of sound signals are numerous and complex, and the tolerance value needs to be set smaller to ensure the accuracy and stability of the decomposition.

[0092] S12, selecting effective IMF components: The IMF components obtained after adaptive VMD decomposition may contain redundant information or noise, so it is necessary to further screen effective IMF components to ensure the accuracy of subsequent feature extraction, including the following steps:

[0093] Calculate the correlation coefficient: Calculate the correlation coefficient between each IMF component and the original signal to evaluate its representativeness to the original signal;

[0094] Set threshold: Set a correlation coefficient threshold based on actual conditions (usually between 0.5 and 0.7, determined according to signal characteristics).

[0095] Screening IMF components: Select IMF components with correlation coefficients greater than the set threshold as valid components. These components are the most relevant to the original signal and retain the main dynamic characteristics.

[0096] S13, Hilbert transform and marginal spectrum sample entropy calculation: The selected effective IMF components are Hilbert transformed to obtain the instantaneous frequency and amplitude envelope of each IMF; then the marginal spectrum sample entropy of these IMF components is calculated to extract the complexity characteristics of the signal.

[0097] Hilbert transform: Perform Hilbert transform on each effective IMF component to obtain the instantaneous amplitude and instantaneous frequency of each component. The instantaneous amplitude reflects the strength of the signal at a certain moment, while the instantaneous frequency indicates the frequency change of the signal at that moment.

[0098] Marginal spectrum calculation: Combine the instantaneous amplitude and instantaneous frequency to obtain the marginal spectrum of each IMF component. The marginal spectrum describes the energy distribution of the signal at different frequencies and is an important manifestation of frequency domain characteristics.

[0099] Sample entropy calculation: Calculate the sample entropy for the marginal spectrum of each IMF component. Sample entropy is an indicator to measure the complexity of the signal. The larger the value, the higher the complexity of the signal.

[0100] S14, constructing a feature vector: combining the sample entropy values ​​of all valid IMF components into a feature vector as the input of the prediction model. This includes the following steps:

[0101] Construction of feature vector: The sample entropy values ​​calculated from the marginal spectrum of each IMF component are arranged in sequence according to the IMF sequence number to form a multidimensional feature vector. The dimension of this vector depends on the number of valid IMFs screened out.

[0102] Normalization: In order to eliminate the scale differences between different features, the feature vectors are normalized so that the numerical range of each feature is consistent, thereby improving the performance of the prediction model.

[0103] It should be noted that the vibration signal and sound signal of the cylinder are both nonlinear, non-stationary and multi-coupled signals, and contain a large amount of noise. Therefore, this embodiment uses adaptive VMD (variational mode decomposition) and Hilbert marginal spectrum sample entropy theory to process the vibration signal and sound signal of the ball mill cylinder, which reduces the impact of noise on signal analysis and helps to identify the dynamic characteristics of the signal.

[0104] The parameter prediction module is used to construct a neural network model to predict the grinding concentration by inputting the vibration signal and grinding sound signal collected in real time, such as Figure 3 As shown, the following steps are included:

[0105] S21. Establish a data set: select test materials, ball stones, and medium filling rate, etc., take grinding concentration as the variable, give different grinding concentrations to represent the load change of the ball mill, collect vibration signals and grinding sound signals under different grinding concentrations, and establish a data set.

[0106] S22, constructing a prediction model: using the BP neural network model to predict the grinding concentration, setting the input layer according to the constructed input feature vector, taking the grinding concentration as the output, and establishing a grinding concentration prediction model, including the following steps:

[0107] S221. Model parameter setting: Set the input layer nodes according to the dimension of the feature vector. The number of input layer nodes is N*M, where N is the number of IMF components and M is the number of features of each component.

[0108] If VMD decomposition results in 5 IMFs and 2 features are extracted from each IMF, then the number of input layer nodes is 5×2=10 nodes.

[0109] The prediction target is grinding concentration, and the number of nodes in the output layer is set to 1.

[0110] The empirical expression is used to calculate the number of neurons in the hidden layer. The formula is:

[0111] ;

[0112] Where: q and p is the total number of neurons in the hidden layer and the output layer.

[0113] S222, model optimization: using a genetic algorithm to optimize the weights and thresholds of the neural network, including the following steps:

[0114] S2221, Initialization: Use the feedforwardnet function to build a 3-layer neural network, determine the weights and thresholds, and express the weights and thresholds in the form of a matrix:

[0115] Construct a weight matrix between the input layer and the hidden layer , represents the connection weight between two layers;

[0116] The threshold matrix of the hidden layer is denoted as , which includes q Row 1 column data;

[0117] Constructing weight matrix between hidden layer and output layer , represents the connection weight between two layers;

[0118] The threshold matrix of the output layer is , which includes a data.

[0119] Where n is the number of nodes in the input layer.

[0120] according to , , , The data in the matrix are encoded according to the arrangement order to form a gene sequence.

[0121] S2222, Adaptability evaluation: Calculate the fitness values ​​of all individuals, and determine the population to continue the next round of iteration based on the value. The higher the fitness value, the greater the probability of being selected. The calculation formula is:

[0122] ;

[0123] In the formula, is the fitness function, x For the gene sequence, K is the grinding concentration amplitude, m is the total number of training data, For sample i The actual grinding concentration, For sample i Enter the predicted values ​​obtained by the model.

[0124] S2223. Perform crossover and mutation operations on all gene sequences according to the designed probabilities. According to the above calculation results, the probability intervals of the crossover and mutation operators are determined to be [0.90, 0.99] and [0.05, 0.15] respectively.

[0125] S2224, input the training data into the network, calculate the error according to the preset error function, and update the weights through the forward and back propagation algorithms, thereby continuously optimizing the parameters and structure of the network to improve the prediction accuracy and generalization ability;

[0126] Use the iterative function to complete the network training and set the learning accuracy. If the network prediction error is less than the learning accuracy, the network training ends and the network weights and thresholds remain unchanged.

[0127] It is understandable that model optimization refers to adjusting network parameters and structures to make the network fit the data better and improve the efficiency and accuracy of life prediction. Since genetic algorithms can be applied to searches in complex spaces and have strong robustness, they have good application effects in scheduling and prediction problems. Therefore, genetic algorithms are used to improve neural networks, and weights and thresholds are optimized through this method.

[0128] S23, parameter prediction: the vibration signal and grinding sound signal obtained by the data acquisition module are processed by the data processing module and then input into the trained prediction model to obtain the grinding concentration.

[0129] The energy consumption calculation module is used to characterize the energy consumption change by using the polynomial regression model through the predicted grinding concentration, set the alarm threshold, and perform online monitoring of the energy consumption change, such as Figure 4 As shown, the following steps are included:

[0130] S31. Data collection and preparation: Collect sufficient historical data, including grinding concentration and corresponding energy consumption values. The data should cover different operating conditions and concentration ranges to ensure the wide applicability of the model; pre-process the data, such as removing outliers and filling missing values, to ensure data quality.

[0131] S32, establish a polynomial regression model: select the model order, and convert energy consumption E Grinding concentration C The relationship is expressed as a polynomial function:

[0132] ;

[0133] In the formula, , ,… are the polynomial regression coefficients, m is the polynomial order.

[0134] It is understandable that low-order models are suitable for simple nonlinear relationships and require less computation. High-order models are suitable for complex nonlinear relationships, but there may be a risk of overfitting, and it is necessary to balance the complexity and generalization ability of the model. Select the best polynomial order through cross-validation. For example, you can try from quadratic to quartic polynomials and select the order that performs best on the validation set.

[0135] S33. Model fitting: Use the least squares method or other optimization algorithms to fit the polynomial model and obtain the polynomial coefficients.

[0136] Specifically, it can be fitted through the PolynomialFeatures function of the programming language library. PolynomialFeatures is a function for generating polynomial features, which is usually used in machine learning and data preprocessing. It is part of the scikit-learn library, specifically located in the sklearn.preprocessing module.

[0137] S34. Model evaluation: Evaluate the model's fitting effect, including residual analysis, R² value (coefficient of determination) and mean square error (MSE) and other indicators. If the model is found to be overfitting, regularization techniques (such as Lasso or Ridge regression) can be used to constrain high-order coefficients to avoid excessive complexity of the model.

[0138] S35. Constructing the energy consumption change trend: inputting the predicted grinding concentration into the polynomial regression model to obtain the energy consumption characterization value, and characterizing the energy consumption change trend by establishing the energy consumption change curve.

[0139] By establishing a polynomial regression model, the energy consumption value is characterized by the grinding concentration, so as to obtain the changing trend of energy consumption. Since there are many factors affecting energy consumption, the error of the energy consumption value calculated indirectly is large. Therefore, this embodiment characterizes the energy consumption by predicting the load state of the mill, and takes the final changing trend as the basis for power saving monitoring, realizing the feedback of energy consumption status, so as to more comprehensively reflect the energy consumption changes in actual operation and provide a basis for power saving optimization.

[0140] The present invention firstly takes the load state of the main controllable factor affecting the energy consumption of the continuous ball mill as the characterization basis of the energy consumption change, monitors the grinding concentration, and then uses the collected vibration signal and grinding sound signal of the cylinder as the basis for judging the grinding concentration. A grinding concentration prediction model is established to obtain the predicted value, and then the predicted value is input into the established polynomial regression model to convert it into the energy consumption change trend, thereby realizing real-time monitoring of the energy consumption of the ball mill, improving the automation degree of the system, and making energy consumption monitoring more efficient.

[0141] In the process of processing collected data, the present invention adopts adaptive VMD and Hilbert marginal spectrum sample entropy theory to construct feature vectors, specifically solves the problems of signal nonlinearity, non-stationarity and multi-coupling characteristics, reduces the influence of noise on signal analysis, and improves the accuracy of grinding concentration prediction; at the same time, in the training of the constructed BP neural network model, the weights and thresholds of the neural network are optimized by adopting a genetic algorithm, thereby ensuring the training effect of the model, avoiding overfitting of the weights and thresholds, improving the convergence speed, and improving the prediction accuracy.

[0142] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A continuous ball mill power saving monitoring system, characterized in that: It includes a data acquisition module, a data processing module, a parameter prediction module and an energy consumption calculation module, wherein the data acquisition module, the data processing module, the parameter prediction module and the energy consumption calculation module are sequentially connected in communication; The data processing module is used to extract features from the collected vibration signals and grinding sound signals and construct feature vectors; the parameter prediction module is used to construct a neural network model to predict the grinding concentration by inputting the real-time collected vibration signals and grinding sound signals; the energy consumption calculation module is used to characterize the energy consumption changes by using a polynomial regression model through the predicted grinding concentration, set an alarm threshold, and monitor the energy consumption changes online; The parameter prediction module predicts the grinding concentration and includes the following steps: Establishing a data set: Select the test material, ball stone and medium filling rate, take grinding concentration as the variable, give different grinding concentrations to represent the load change of the ball mill, collect the vibration signal and grinding sound signal under different grinding concentrations, and establish a data set; Construct a prediction model: Use the BP neural network model to predict the grinding concentration, set the input layer according to the constructed input feature vector, take the grinding concentration as the output, and establish a grinding concentration prediction model; Parameter prediction: The vibration signal and grinding sound signal obtained by the data acquisition module are processed by the data processing module and then input into the trained prediction model to obtain the grinding concentration; The data processing module constructs a feature vector including the following steps: Adaptive VMD algorithm decomposition: The collected cylinder vibration signal and grinding sound signal are decomposed by the adaptive VMD algorithm to obtain their respective IMF components; Selecting effective IMF components: The IMF components obtained after adaptive VMD decomposition are further screened by using the correlation coefficient method to obtain effective IMF components; Hilbert transform and marginal spectrum sample entropy calculation: The selected effective IMF components are Hilbert transformed to obtain the instantaneous frequency and amplitude envelope of each IMF; then the marginal spectrum sample entropy of each IMF component is calculated; Construct feature vector: combine the sample entropy values ​​of all valid IMF components into a feature vector as the input of the prediction model; The adaptive VMD algorithm decomposition comprises the following steps: Initialization parameters: Set the initial mode number K, penalty parameter α and convergence tolerance of the vibration signal and grinding sound signal decomposition respectively ; Iterative update: decompose the signal into several modes with limited frequency bands, so that the center frequency and bandwidth of each mode converge to a stable state; Decomposition results: After the iteration is completed, the IMF components of the vibration signal and the grinding sound signal are obtained. Each component reflects the dynamic characteristics of different frequency components in the signal. The selecting of effective IMF components comprises the following steps: Calculate the correlation coefficient: Calculate the correlation coefficient between each IMF component and the original signal to evaluate its representativeness to the original signal; Set threshold: Set a correlation coefficient threshold according to the actual situation; Screening IMF components: Select IMF components with correlation coefficients greater than the set threshold as valid components; The Hilbert transform and marginal spectrum sample entropy calculation include the following steps: Hilbert transform: Perform Hilbert transform on each effective IMF component to obtain the instantaneous amplitude and instantaneous frequency of each component; Marginal spectrum calculation: Combine the instantaneous amplitude and instantaneous frequency to obtain the marginal spectrum of each IMF component; Sample entropy calculation: Calculate the sample entropy for the marginal spectrum of each IMF component; The constructing of the feature vector comprises the steps of: Construction of feature vector: The sample entropy values ​​calculated from the marginal spectrum of each IMF component are arranged in sequence according to the IMF sequence number to form a multi-dimensional feature vector. The dimension of the vector depends on the number of valid IMFs screened out. Normalization: Normalize the feature vectors to make the value range of each feature consistent; The construction of the prediction model comprises the following steps: Model parameter setting: Set the input layer nodes according to the dimension of the feature vector, and the number of input layer nodes is N M, where N is the number of IMF components and M is the number of features of each component; The prediction target is grinding concentration, and the number of nodes in the output layer is set to 1; The empirical expression is used to calculate the number of neurons in the hidden layer. The formula is: ； Where: and is the total number of neurons in the hidden layer and output layer; The construction of the prediction model also includes the following steps: Model optimization: Genetic algorithm is used to optimize the weights and thresholds of the neural network, including the following steps: Initialization: Use the feedforwardnet function to build a 3-layer neural network, determine the weights and thresholds, and express the weights and thresholds in the form of a matrix: Construct a weight matrix between the input layer and the hidden layer , represents the connection weight between two layers; The threshold matrix of the hidden layer is denoted as , which includes q Row 1 column data; Constructing weight matrix between hidden layer and output layer , represents the connection weight between two layers; The threshold matrix of the output layer is , which includes a data; in, is the number of nodes in the input layer; according to , , , The data in the matrix are encoded in the order of arrangement to form a gene sequence; Adaptability evaluation: Calculate the fitness values ​​of all individuals according to the fitness function, and determine the population to continue the next round of iteration based on the fitness values. The higher the fitness value, the greater the probability of being selected. Perform crossover and mutation operations on all gene sequences according to the designed probabilities. According to the calculation results, the probability intervals of crossover and mutation operators are determined to be [0.90, 0.99] and [0.05, 0.15] respectively; Input the training data into the network, calculate the error according to the preset error function, update the weights through the forward and back propagation algorithms, and continuously optimize the parameters and structure of the network; Use the iterative function to complete the network training and set the learning accuracy. When the network prediction error is less than the learning accuracy, the network training ends and the network weights and thresholds remain unchanged. The fitness function is calculated as follows: ； In the formula, is the fitness function, x For the gene sequence, K is the grinding concentration amplitude, m is the total number of training data, For sample The actual grinding concentration, For sample Enter the predicted values ​​obtained by the model; The energy consumption calculation module characterizes the energy consumption change and includes the following steps: Data collection and preparation: Collect historical data, including grinding concentration and corresponding energy consumption values. The data covers different working conditions and concentration ranges, and pre-process the data, including removing outliers and filling missing values; Establish a polynomial regression model: select the model order and convert energy consumption E Grinding concentration C The relationship is expressed as a polynomial function: ； In the formula, , , are the polynomial regression coefficients, is the polynomial order; Model fitting: Use the PolynomialFeatures function to fit the polynomial model and obtain the polynomial coefficients; The characterization of energy consumption changes further includes the steps of: Model evaluation: Evaluate the model's fitting effect, including residual analysis, R² value and mean square error. When the model is overfitted, regularization techniques are used to constrain high-order coefficients. Constructing energy consumption change trend: input the predicted grinding concentration into the polynomial regression model to obtain the energy consumption characterization value, and characterize the energy consumption change trend by establishing the energy consumption change curve; The data acquisition module is used to collect vibration signals and grinding sound signals in real time, and transmit the collected signals to the data processing module through the data acquisition card, wherein the vibration signal is collected by the acceleration sensor, and the grinding sound signal is collected by the microphone.

Images