Torque prediction method suitable for high-viscosity material mixing
By using an artificial neural network model to predict torque during the mixing process of high-viscosity materials, the problems of insufficient accuracy and stability of traditional torque measurement methods under high temperature and high pressure environments are solved. Torque prediction during the mixing process of high-viscosity materials is realized, improving the accuracy of measurement and the operating efficiency of equipment.
Patent Information
- Application Number
- CN202511347925.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-19
- Publication Date
- 2026-02-10
AI Technical Summary
Existing technologies for torque measurement in the mixing process of high-viscosity and elasto-plastic materials suffer from problems such as limited sensor installation space, poor stability, and insufficient accuracy, especially in high-temperature, high-pressure, or corrosive environments where accurate measurement is difficult to achieve.
By collecting data on rotational speed, pressure drop, temperature, and flow rate, torque prediction is performed using an artificial neural network model. The trained model is transferable and applicable to various equipment and environments, enabling rapid and accurate torque prediction.
In situations where torque sensors cannot be installed, torque prediction during the mixing process of high-viscosity materials is achieved, improving the accuracy and stability of the measurement, adapting to different environments, and reducing maintenance costs and downtime.
Smart Images
Figure CN121503191A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a torque prediction method applicable to the mixing of high-viscosity materials, belonging to the field of hydraulic equipment. Background Technology
[0002] In key industrial sectors such as chemical, pharmaceutical, and food processing, hydraulic mixing equipment plays a crucial role in the mixing of high-viscosity and elasto-plastic materials. Accurate torque measurement is essential for ensuring the uniformity and efficiency of the mixing process, directly impacting product quality and production safety. Currently, mainstream torque measurement methods include using torque sensors and estimating torque using formulas. However, the application of torque sensors is limited in certain scenarios. For example, the tight integration of the hydraulic motor with the working parts restricts the sensor's installation space; in high-temperature, high-pressure, or corrosive environments, the stability and durability of sensors also face challenges. In situations where direct torque measurement with sensors is not possible, traditional torque estimation methods, including energy conversion, balanced force, and transmission methods, are limited by insufficient accuracy or unsuitability for specific operating conditions. Traditional torque estimation methods have limitations in capturing dynamic signal changes, installation complexity, cost, and adaptability to specific environments. Therefore, new torque measurement methods are urgently needed for hydraulic mixing equipment handling high-viscosity and elasto-plastic materials. Summary of the Invention
[0003] To address the torque measurement problem in mixing high-viscosity and elasto-plastic materials, the present invention aims to provide a torque prediction method suitable for mixing high-viscosity materials. By collecting data on rotational speed, pressure drop, temperature, and flow rate, an artificial neural network model is trained using existing torque data. The trained neural network model is transferable and applicable to various equipment and environments, enabling rapid and accurate prediction of torque in hydraulic mixing equipment.
[0004] The objective of this invention is achieved through the following technical solution:
[0005] This invention discloses a torque prediction method suitable for mixing high-viscosity materials. It utilizes rotational speed, pressure drop in hydraulic power transmission, temperature, and flow rate as inputs, and the actual torque measured by a torque sensor as the output. By training an artificial neural network model, it achieves accurate torque prediction for hydraulic mixing equipment. Through machine learning, the artificial neural network model is continuously optimized to improve prediction accuracy. The artificial neural network model has good transferability, providing accurate torque prediction results for operating conditions where torque sensors cannot be installed. The specific steps include:
[0006] Step 1: Sensor installation and data acquisition system setup;
[0007] Install sensors according to the equipment structure and measurement requirements, connect signal lines to the data acquisition card and complete initialization to ensure accurate and stable data acquisition;
[0008] Step 2: Set up multiple operating scenarios, collect torque values, and establish a raw dataset;
[0009] Construct multiple operating conditions, considering material type, viscosity, mixing ratio and working time, record parameters and operating data before and after equipment start-up, collect torque values as training labels, and establish the original dataset.
[0010] Step 3: Preprocessing of the original dataset, including the following sub-steps:
[0011] Step 3.1 Filter key features and eliminate redundant features;
[0012] As a preferred method, the Pearson correlation coefficient is used to assess the importance of features, screen key features, and eliminate redundant features, as shown in Equation (1):
[0013]
[0014] Where, r pq p is the Pearson correlation coefficient. i q i Let be the i-th observations of variables p and q, respectively. are the means of variables p and q, respectively, and n is the sample size;
[0015] By using the correlation coefficient, we can initially determine the degree of linear correlation between each variable and the torque value. The closer the absolute value of the correlation coefficient is to 1, the stronger the linear correlation.
[0016] Step 3.2 Remove outliers;
[0017] As a preferred method, the K-nearest neighbor algorithm is used to remove outliers. Anomalies are judged based on the K nearest neighbors of a data point, and the distance between each point and its nearest neighbor is determined. The one-dimensional K-nearest neighbor algorithm is shown in equation (2):
[0018] d(x a ,x b )=|x a -x b | (2)
[0019] Where d(x) a ,x b ) is x a x b The distance between them, x a x b If two adjacent data points in the dataset are far apart, they are considered abnormal and removed.
[0020] Step 3.3 Eliminate the influence of dimensions;
[0021] The minimax normalization method is used to eliminate the influence of dimensions, as shown in equation (3):
[0022]
[0023] Where, x norm It is the normalized value obtained after min-max normalization, x min x max These are the minimum and maximum values of variable x in the dataset, respectively.
[0024] The min-max normalization method is used to eliminate the influence of different dimensions between variables, so that the data have the same weight in the training of the neural network.
[0025] Step 4: Build an artificial neural network and construct a torque prediction model, including the following sub-steps:
[0026] Step 4.1 Determine the number of input layer nodes based on the number of input variables after filtering;
[0027] After data correlation analysis and filtering, m variables are retained, and the input layer is set with m nodes;
[0028] Step 4.2 Determine the number of nodes in the hidden layer through multiple tests and optimizations;
[0029] A well-designed hidden layer can fully explore the complex nonlinear relationships between input variables and improve the model's predictive ability.
[0030] Step 4.3 Determine the output layer nodes;
[0031] The output layer nodes are used to output the predicted torque value. The output layer does not use an activation function and directly outputs the predicted value.
[0032] Step 4.4 Set the activation function for the hidden layer;
[0033] Preferably, the neurons in the hidden layer use the ReLU (Rectified Linear Unit) activation function, as shown in equation (4):
[0034] f(z) = max(0,z) (4)
[0035] Where f(z) is the ReLU activation function, and z is the model input value;
[0036] The ReLU function is simple to compute, converges quickly, and can introduce nonlinear characteristics to enhance the expressive power of the model.
[0037] Step 4.5 Determine the training set, validation set, and test set for the torque prediction model;
[0038] The training set, validation set, and test set are determined based on the preprocessed dataset and input features.
[0039] The training set is used for training, enabling the torque prediction model to learn the mapping relationship between input variables and torque;
[0040] The validation set is used for testing after each training round to prevent overfitting during training.
[0041] The test set is used to evaluate performance and verify the accuracy of the torque prediction model on unseen data;
[0042] Step 4.6 Determine the training parameters for the torque prediction model;
[0043] As a preferred approach, the Adam algorithm is used to iteratively update the parameters of the torque prediction model, thereby gradually reducing the loss function value, as shown in equation (5):
[0044]
[0045] Where, θ t+1 Let θ be the torque prediction model parameters updated after the (t+1)th iteration. t Here are the torque prediction model parameters at the t-th iteration, α is the learning rate, and m... t For the corrected first-order moment estimate, v t This is the corrected second-order moment estimate, where ∈ is a small constant used to ensure numerical stability;
[0046] Setting the learning rate α controls the step size of the torque prediction model parameter update. If the learning rate is too large, the torque prediction model will not converge. If the learning rate is too small, the torque prediction model will be trained too slowly. As a preferred option, the learning rate α is set based on experience.
[0047] Determine the loss function, assess the degree of difference between the torque prediction model's predicted values and the actual values, and evaluate the performance of the torque prediction model;
[0048] Preferably, the loss function includes mean squared error and cross-entropy loss;
[0049] When optimizing the model, the loss function value is minimized by continuously adjusting the torque prediction model parameters, and the prediction results of the torque prediction model are closer to the real situation.
[0050] Step 5: Train the torque prediction model using a feedforward neural network, including the following sub-steps:
[0051] Step 5.1 Input the training data into the torque prediction model. The data starts from the input layer, passes through the hidden layers, and finally reaches the output layer.
[0052] In each neuron, a weighted summation and activation function are performed. For the k-th neuron in the j-th hidden layer, its input is as shown in equation (6):
[0053] z jk =∑ i w j,k,i x j-1,i +b j,k (6)
[0054] Among them, z jk For the input of the k-th neuron in the j-th layer, w j,k,i The weights connecting the i-th neuron in layer (j-1) and the k-th neuron in layer j are given by x. j-1,i b is the output of the i-th neuron in the (j-1)-th layer. j,k This is the bias of the k-th neuron in the j-th layer;
[0055] The output of a neuron is determined by an activation function, and the predicted torque value is finally obtained in the output layer.
[0056] Step 5.2: Based on the predicted values of the torque prediction model As a preferred method, the loss value l is determined using the mean square error loss function, as shown in equation (7): (The actual torque value y is used in conjunction with the actual torque value y.)
[0057]
[0058] The loss value reflects the degree of difference between the torque prediction model's prediction results and the actual values. The goal is to minimize the loss value by adjusting the parameters of the torque prediction model.
[0059] Step 5.3: Based on the loss value, update the weights and biases of the torque prediction model using the backpropagation algorithm;
[0060] The backpropagation algorithm uses the chain rule to determine the gradient of the loss function with respect to each weight and bias;
[0061] For weight w j,k,i The gradient is shown in equation (8):
[0062]
[0063] Where L is the loss function, w j,k,i Let z be the connection weight between the i-th neuron in the (j-1)-th layer and the k-th neuron in the j-th layer of the neural network. j,k This is the input to the k-th neuron in the j-th layer;
[0064] For bias b j,k The gradient is shown in equation (9):
[0065]
[0066] Among them, b j,k Let z be the bias of the k-th neuron in the j-th layer, and z is a learnable parameter in the neural network. j,k This is the input value of the k-th neuron in the j-th layer;
[0067] According to the weight w j,k,i gradient and bias b j,k The gradient is used to iteratively update the weights and biases using an optimizer until the torque prediction model converges.
[0068] Step 5.4: Evaluate the performance of the torque prediction model.
[0069] As a preferred option, the standard mean square error (MSE) is used as the evaluation index for the torque prediction model to measure the average error between the predicted and actual values of the torque prediction model, as shown in equation (10):
[0070]
[0071] Where n is the number of samples, y i This is the actual torque value. The torque value predicted by the torque prediction model;
[0072] Using the coefficient of determination R 2 The goodness of fit of the regression model is measured as shown in Equation (11):
[0073]
[0074] Among them, y i For the true value, For predicted values, The mean of the true values;
[0075] Coefficient of determination R 2 The closer the value is to 1, the better the torque prediction model fits the data, meaning the higher the proportion of variability of the dependent variable that the torque prediction model can explain.
[0076] It also includes step six: Based on the torque prediction model obtained in step five, it can accurately predict the torque value of the hybrid equipment under different working conditions. It is applicable to torque measurement scenarios where torque sensors cannot be installed or where the measuring instruments may be damaged, thus solving related engineering problems.
[0077] Beneficial effects:
[0078] 1. The present invention provides a torque prediction method suitable for mixing high-viscosity materials. It selects rotational speed, pressure drop, temperature and flow rate as input parameters. The parameters are easy to measure, which solves the problem of traditional torque sensor installation. Even in a small space where the hydraulic motor and working parts are tightly connected, or in harsh environments such as high temperature, high pressure and strong corrosion, it can still collect data stably and is not limited by torque sensor installation conditions and environmental factors.
[0079] 2. The present invention provides a torque prediction method applicable to the mixing of high-viscosity materials. Through multi-condition testing, it comprehensively collects data under different material types, viscosities, mixing ratios, and working durations, covering various working scenarios of hydraulic mixing equipment. The data learned by the model is highly representative. Pearson correlation coefficient is used to screen key features, effectively removing redundant information. Normalization is applied to the data to improve the training effect. The feedforward neural network is continuously optimized, the number of nodes in each layer is reasonably set, and the ReLU activation function is selected to enhance the model's expressive power. The model is trained by combining the Adam algorithm and the mean squared error loss function, so that the model approaches the optimal solution in continuous iteration, achieving high-precision torque prediction.
[0080] 3. The present invention provides a torque prediction method applicable to the mixing of high-viscosity materials. The trained artificial neural network model has good transferability and can adapt to different hydraulic mixing equipment. It can quickly achieve high-precision torque prediction without complicated debugging, allowing operators to grasp the equipment operating status in real time and accurately, promptly identify potential problems and adjust production parameters, effectively avoid hydraulic mixing equipment failures caused by abnormal torque, reduce maintenance costs and downtime, and improve production efficiency. Attached Figure Description
[0081] Figure 1 This is a flowchart of a torque prediction method applicable to the mixing of high-viscosity materials according to the present invention;
[0082] Figure 2 This is a schematic diagram of the artificial neural network model structure in the embodiment;
[0083] Figure 3 This is a schematic diagram of the artificial neural network training process in the embodiment;
[0084] Figure 4 This is a comparison chart of the actual and predicted values of the torque of the hydraulic mixing equipment in the embodiment. Detailed Implementation
[0085] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.
[0086] Example 1:
[0087] This embodiment uses a 5L vertical mixing device as an example, and applies a torque prediction method suitable for mixing high-viscosity materials according to the present invention to predict the torque of mixing high-viscosity materials, such as... Figure 1 As shown, it includes the following steps:
[0088] Step 1: Sensor installation and data acquisition system setup;
[0089] Install sensors according to the equipment structure and measurement requirements, connect signal lines to the data acquisition card and complete initialization to ensure accurate and stable data acquisition;
[0090] In this embodiment, flow sensors are installed on suitable straight sections of the hydraulic motor's inlet and outlet oil circuits, and the signal lines are connected to a data acquisition card.
[0091] Pressure sensors are installed at the inlet and outlet pipes of the hydraulic motor, and the signal lines are connected to the data acquisition card.
[0092] Install thermistor temperature sensors at the inlet and outlet pipes of the hydraulic motor oil circuit, and connect the signal lines to the data acquisition card.
[0093] Install a speed sensor on the rotating shaft of the mixing equipment, ensuring that the sensor's installation position is aligned with the markings on the rotating shaft, and connect the signal line to the data acquisition card;
[0094] A torque sensor is installed on the rotating shaft of the mixing equipment, and the signal line is connected to a data acquisition card.
[0095] The sensor types and parameters involved in this example are shown in Table 1:
[0096] Table 1 Key Information of the Main Sensors Selected
[0097] Serial Number Sensor type Measurement points Key information 1 Resistance Temperature Detector Hydraulic motor oil inlet PT100, -40~100℃ 2 Resistance Temperature Detector Hydraulic motor oil circuit outlet PT100, -40~100℃ 3 pressure transmitter Hydraulic motor oil inlet HPTM180, 0~16MPa, temperature and pressure integrated 4 pressure transmitter Hydraulic motor oil circuit outlet HPTM180, 0~16MPa, temperature and pressure integrated 5 Turbine flow meter Hydraulic motor oil inlet <![CDATA[LWGB-10,0~1.2m 3 / H]]> 6 Turbine flow meter Hydraulic motor oil circuit outlet <![CDATA[LWGB-10,0~1.2m 3 / H]]> 7 Speed sensor Hydraulic motor drive shaft Hall effect, square wave output, 0–20kHz 8 Torque sensor Below the hydraulic motor Custom design, 0~200Nm
[0098] After connecting all sensor signal lines to the data acquisition card, configure the parameters of the data acquisition card.
[0099] Based on the fact that the torque value update frequency does not need to be too high in actual production, and considering the memory and computing power, a low sampling frequency is set. In this example, the sampling rate is set to 1Hz. The data acquisition card is connected to the computer via USB, the corresponding data acquisition software is installed, and the system initialization settings are performed.
[0100] Step 2: Set up multiple operating scenarios, collect torque values, and establish a raw dataset;
[0101] Construct multiple operating conditions, considering material type, viscosity, mixing ratio and working time, record parameters and operating data before and after equipment start-up, collect torque values as training labels, and establish the original dataset.
[0102] This example simulates mixtures of different viscosities by mixing fire-retardant coatings with different formulations, and sets different rotation speeds and stirring times to obtain measurement values under different working conditions. All collected data is stored in a computer database and managed using a MySQL database.
[0103] Step 3: Preprocessing of the original dataset, including the following sub-steps:
[0104] Step 3.1 Filter key features and eliminate redundant features;
[0105] In this embodiment, the Pearson correlation coefficient is used to assess feature importance, screen key features, and eliminate redundant features, as shown in Equation (1):
[0106]
[0107] Where, r pq p is the Pearson correlation coefficient. i q i Let be the i-th observations of variables p and q, respectively. are the means of variables p and q, respectively, and n is the sample size;
[0108] By using the correlation coefficient, we can initially determine the degree of linear correlation between each variable and the torque value. The closer the absolute value of the correlation coefficient is to 1, the stronger the linear correlation.
[0109] Based on the Pearson correlation coefficient, features that are strongly correlated with torque values and are of high importance for torque prediction are selected.
[0110] In this embodiment, based on the measurement results, the correlation is ranked from strongest to weakest as follows: inlet pressure > outlet pressure > inlet flow rate > outlet flow rate > rotational speed > inlet temperature > outlet temperature;
[0111] Step 3.2 Remove outliers;
[0112] In this embodiment, the K-nearest neighbor algorithm is used to remove outliers. Anomalies are judged based on the K nearest neighbors of a data point, and the distance between each point and its nearest neighbor is determined. The one-dimensional K-nearest neighbor algorithm is shown in equation (2):
[0113] d(x a ,x b )=|x a -x b | (2)
[0114] Where d(x) a ,x b ) is x a x b The distance between them, x a xb For two adjacent data points in the dataset;
[0115] If the distance exceeds the threshold, it is considered abnormal and removed.
[0116] Step 3.3 Eliminate the influence of dimensions;
[0117] The minimax normalization method is used to eliminate the influence of dimensions, as shown in equation (3):
[0118]
[0119] Where, x norm It is the normalized value obtained after min-max normalization, x min x max These are the minimum and maximum values of variable x in the dataset, respectively.
[0120] The min-max normalization method is used to eliminate the influence of different dimensions between variables, so that the data have the same weight in the training of the neural network.
[0121] Step 4: Build an artificial neural network and construct a torque prediction model, including the following sub-steps:
[0122] In this embodiment, the artificial neural network is a feedforward neural network (FFNet), with the following structure: Figure 2 As shown, the first layer is the input layer, the last layer is the output layer, the hidden layers in the middle are not connected to the outside, the nodes of each layer are connected to the nodes of the next layer, there are no connections between nodes in the same layer, and the input data only passes through all nodes once before entering the output layer.
[0123] Step 4.1 Determine the number of input layer nodes based on the number of input variables after filtering;
[0124] After data correlation analysis and filtering, m variables are retained, and the input layer is set with m nodes;
[0125] In this embodiment, the input layer retains 7 variables and sets 7 nodes;
[0126] Step 4.2 Determine the number of nodes in the hidden layer through multiple tests and optimizations;
[0127] A well-designed hidden layer can fully explore the complex nonlinear relationships between input variables and improve the model's predictive ability.
[0128] In this embodiment, there are two hidden layers, with 200 nodes in the first layer and 100 nodes in the second layer.
[0129] Step 4.3 Determine the output layer nodes;
[0130] The output layer nodes are used to output the predicted torque value. The output layer does not use an activation function and directly outputs the predicted value.
[0131] Step 4.4: Identify the neurons in the hidden layer;
[0132] In this embodiment, the neurons in the hidden layer use the ReLU (Rectified Linear Unit) activation function, as shown in equation (4):
[0133] f(z) = max(0,z) (4)
[0134] Where f(z) is the ReLU activation function, and z is the model input value;
[0135] The ReLU function is simple to compute, converges quickly, and can introduce nonlinear characteristics to enhance the expressive power of the model.
[0136] Step 4.5 Determine the training set, validation set, and test set for the torque prediction model;
[0137] The training set, validation set, and test set are determined based on the preprocessed dataset and input features.
[0138] The training set is used for training, enabling the model to learn the mapping relationship between input variables and torque;
[0139] The validation set is used for testing after each training round to prevent overfitting during training.
[0140] The test set is used to evaluate performance and verify the accuracy of the torque prediction model on unseen data;
[0141] In this embodiment, after data preprocessing, a total of 11,602 samples were obtained, and the training set, validation set, and test set were divided in a ratio of 7:1.5:1.5.
[0142] Step 4.6 Determine the training parameters for the torque prediction model;
[0143] In this embodiment, the Adam algorithm is used to update the parameters of the torque prediction model iteratively, so that the loss function value gradually decreases, as shown in equation (5):
[0144]
[0145] Where, θ t+1 Let θ be the torque prediction model parameters updated after the (t+1)th iteration. t Here are the torque prediction model parameters at the t-th iteration, α is the learning rate, and m... t For the corrected first-order moment estimate, v t This is the corrected second-order moment estimate, where ∩ is a small constant used to ensure numerical stability;
[0146] Setting the learning rate α controls the step size of the torque prediction model parameter update. If the learning rate is too large, the torque prediction model will not converge, and if the learning rate is too small, the torque prediction model will be trained too slowly.
[0147] In this embodiment, the learning rate α is set to 0.001 based on experience.
[0148] Determine the loss function, assess the degree of difference between the torque prediction model's predicted values and the actual values, and evaluate the performance of the torque prediction model;
[0149] In this embodiment, the loss function includes mean squared error and cross-entropy loss;
[0150] When optimizing the model, the loss function value is minimized by continuously adjusting the torque prediction model parameters, and the prediction results of the torque prediction model are closer to the real situation.
[0151] Step 5: Train the torque prediction model using a feedforward neural network, including the following sub-steps:
[0152] Step 5.1 Input the training data into the torque prediction model. The data starts from the input layer, passes through the hidden layers, and finally reaches the output layer.
[0153] In each neuron, a weighted summation and activation function are performed. For the k-th neuron in the j-th hidden layer, its input is as shown in equation (6):
[0154] z jk =∑ i w j,k,i x j-1,i +b j,k (6)
[0155] Among them, z jk For the input of the k-th neuron in the j-th layer, w j,k,i The weights connecting the i-th neuron in layer (j-1) and the k-th neuron in layer j are given by x. j-1,i b is the output of the i-th neuron in the (j-1)-th layer. j,k This is the bias of the k-th neuron in the j-th layer;
[0156] The output of a neuron is determined by an activation function, and the predicted torque value is finally obtained in the output layer.
[0157] Step 5.2: Based on the predicted values of the torque prediction model The loss value l is determined by comparing it with the actual torque value y;
[0158] In this embodiment, the loss value l is determined using the mean square error loss function, as shown in equation (7):
[0159]
[0160] The loss value reflects the degree of difference between the torque prediction model's prediction results and the actual values. The goal is to minimize the loss value by adjusting the parameters of the torque prediction model.
[0161] Step 5.3: Based on the loss value, update the weights and biases of the torque prediction model using the backpropagation algorithm;
[0162] The backpropagation algorithm uses the chain rule to determine the gradient of the loss function with respect to each weight and bias;
[0163] For weight w j,k,i The gradient is shown in equation (8):
[0164]
[0165] Where L is the loss function, w j,k,i Let z be the connection weight between the i-th neuron in the (j-1)-th layer and the k-th neuron in the j-th layer of the neural network. j,k This is the input to the k-th neuron in the j-th layer;
[0166] For bias b j,k The gradient is shown in equation (9):
[0167]
[0168] Among them, b j,k Let z be the bias of the k-th neuron in the j-th layer, and z is a learnable parameter in the neural network. j,k This is the input value of the k-th neuron in the j-th layer;
[0169] According to the weight w j,k,i gradient and bias b j,k The gradient is used to iteratively update the weights and biases using an optimizer until the torque prediction model converges.
[0170] Step 5.4: Torque prediction model evaluation;
[0171] In this embodiment, the standard mean square error (MSE) is used as the evaluation index for the torque prediction model to measure the average error between the predicted value and the actual value, as shown in Equation (10):
[0172]
[0173] Where n is the number of samples, y i This is the actual torque value. The torque value predicted by the torque prediction model;
[0174] Using the coefficient of determination R 2 The goodness of fit of the regression model is measured as shown in Equation (11):
[0175]
[0176] Among them, y i For the true value, For predicted values, The mean of the true values;
[0177] Coefficient of determination R 2 The closer the value is to 1, the better the torque prediction model fits the data, meaning the higher the proportion of variability of the dependent variable that the torque prediction model can explain.
[0178] In this embodiment, the training process of the torque prediction model is as follows: Figure 3 As shown, the dynamic changes of the entire training process are displayed in detail. The error curves of the training set and the validation set have always maintained a relatively reasonable gap. There was no significant increase in the validation set error in the later stage of training, indicating that there was no overfitting.
[0179] In the embodiment, the comparison between the actual value and the predicted value is as follows: Figure 4 As shown, the sample points are basically concentrated on the diagonal, and there is no dispersion of outliers.
[0180] To more objectively and accurately evaluate the performance of the established torque prediction model, an independent test set was used to test the model. The mean square error (MSE) of the torque prediction model established in this embodiment was found to be 0.236. The mean square error measures the average of the squares of the errors between the predicted and actual values. The smaller this value, the closer the predicted values of the torque prediction model are to the actual values, and the higher the prediction accuracy of the model. The coefficient of determination R0 of the torque prediction model is... 2 The coefficient of determination is 0.978. It reflects the goodness of fit of the torque prediction model to the data. The closer the coefficient of determination is to 1, the higher the proportion of the change in the dependent variable that the torque prediction model can explain, and the better the torque prediction model fits the data.
[0181] Combining MSE and R 2 These two key metrics demonstrate that the artificial neural network model established using the method of this invention exhibits excellent performance and high reliability in this training and testing.
[0182] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A torque prediction method suitable for mixing high-viscosity materials, characterized in that: Using speed, pressure drop, temperature, and flow rate in hydraulic power transmission as inputs, and the torque actually measured by the torque sensor as the output, an artificial neural network model is trained to accurately predict the torque of hydraulic mixing equipment. Through machine learning, the artificial neural network model is continuously optimized to improve the accuracy of the prediction. The artificial neural network model can provide accurate torque prediction results for working conditions where torque sensors cannot be installed.
2. The torque prediction method for high-viscosity material mixing as described in claim 1, characterized in that: Specifically, the steps include the following: Step 1: Sensor installation and data acquisition system setup; Install the sensors according to the equipment structure and measurement requirements, connect the signal lines to the data acquisition card and complete the initialization; Step 2: Set up multiple operating scenarios, collect torque values, and establish a raw dataset; Construct multiple operating conditions, considering material type, viscosity, mixing ratio and working time, record parameters and operating data before and after equipment start-up, collect torque values as training labels, and establish the original dataset. Step 3: Preprocessing of the original dataset, including the following sub-steps: Step 3.1 Filter key features and eliminate redundant features; Step 3.2 Remove outliers; Step 3.3 Eliminate the influence of dimensions; Step 4: Build an artificial neural network and construct a torque prediction model, including the following sub-steps: Step 4.1 Determine the number of input layer nodes based on the number of input variables after filtering; After data correlation analysis and filtering, m variables are retained, and the input layer is set with m nodes; Step 4.2 Determine the number of nodes in the hidden layer through multiple tests and optimizations; Step 4.3 Determine the output layer nodes; Step 4.4: Set the activation function for the hidden layer; Step 4.5 Determine the training set, validation set, and test set for the torque prediction model; The training set, validation set, and test set are determined based on the preprocessed dataset and input features. The training set is used to train the torque prediction model and build a mapping relationship between input variables and torque; The validation set is used for testing after each training round to prevent overfitting during training. The test set is used to evaluate performance and verify the accuracy of the torque prediction model on unseen data; Step 4.6 Determine the training parameters for the torque prediction model; Determine the loss function, assess the degree of difference between the torque prediction model's predicted values and the actual values, and evaluate the performance of the torque prediction model; When optimizing the model, the loss function value is minimized by continuously adjusting the torque prediction model parameters, and the prediction results of the torque prediction model are closer to the real situation. Step 5: Train the torque prediction model using a feedforward neural network, including the following sub-steps: Step 5.1 Input the training data into the torque prediction model. The data starts from the input layer, passes through the hidden layers, and finally reaches the output layer. In each neuron, a weighted summation and activation function are performed. For the k-th neuron in the j-th hidden layer, its input is as shown in equation (1): z jk =∑ i w j,k,i x j-1,i +b j,k (1) Among them, z jk For the input of the k-th neuron in the j-th layer, w j,k,i The weights connecting the i-th neuron in layer (j-1) and the k-th neuron in layer j are given by x. j-1,i b is the output of the i-th neuron in the (j-1)-th layer. j,k This is the bias of the k-th neuron in the j-th layer; The output of the neuron is determined by the activation function, and the torque prediction value is finally obtained in the output layer. Step 5.2: Based on the predicted values of the torque prediction model The loss value is determined by comparing it with the actual torque value y; The loss value reflects the degree of difference between the torque prediction model's prediction results and the actual values. The goal is to minimize the loss value by adjusting the parameters of the torque prediction model. Step 5.3: Based on the loss value, update the weights and biases of the torque prediction model using the backpropagation algorithm; The backpropagation algorithm uses the chain rule to determine the gradient of the loss function with respect to each weight and bias; For weight w j,k,i The gradient is shown in equation (2): Where L is the loss function, z j,k This is the input to the k-th neuron in the j-th layer; For bias b j,k The gradient is shown in equation (3): According to the weight w j,k,i gradient and bias b j,k The gradient is used to iteratively update the weights and biases using an optimizer until the torque prediction model converges. Step 5.4: Evaluate the performance of the torque prediction model.
3. The torque prediction method for high-viscosity material mixing as described in claim 2, characterized in that: In step 3.1, the Pearson correlation coefficient is used to assess the importance of features, screen key features, and eliminate redundant features, as shown in equation (4): Where, r pq p is the Pearson correlation coefficient. i q i Let be the i-th observation values of variables p and q, respectively. are the means of variables p and q, respectively, and n is the sample size; By using the correlation coefficient, we can initially determine the degree of linear correlation between each variable and the torque value. The closer the absolute value of the correlation coefficient is to 1, the stronger the linear correlation.
4. The torque prediction method for high-viscosity material mixing as described in claim 2, characterized in that: In step 3.2, the K-nearest neighbor algorithm is used to remove outliers. Anomalies are judged based on the K nearest neighbors of the data points, and the distance between each point and its nearest neighbor is determined. The one-dimensional K-nearest neighbor algorithm is shown in equation (5): d(x a ,x b )=|x a -x b | (5) Where d(x) a ,x b ) is x a x b The distance between them, x a x b For two adjacent data points in the dataset; If the distance exceeds the threshold, it is considered abnormal and removed.
5. The torque prediction method for high-viscosity material mixing as described in claim 2, characterized in that: In step 3.3, the minimax normalization method is used to eliminate the influence of dimensions, as shown in equation (6): Where, x norm It is the normalized value obtained after min-max normalization, x min x max These are the minimum and maximum values of variable x in the dataset, respectively. The min-max normalization method eliminates the influence of different dimensions between variables, so that the data have the same weight in the training of the neural network.
6. The torque prediction method for high-viscosity material mixing as described in claim 2, characterized in that: In step 4.4, the neurons in the hidden layer are activated using the ReLU activation function, as shown in equation (7): f(z) = max(0,z) (7) Where f(z) is the ReLU activation function and z is the model input value.
7. The torque prediction method for high-viscosity material mixing as described in claim 2, characterized in that: In step 4.6, the Adam algorithm is used to iteratively update the parameters of the torque prediction model, so that the loss function value gradually decreases, as shown in equation (8): Where, θ t+1 Let θ be the torque prediction model parameters updated after the (t+1)th iteration. t Here are the torque prediction model parameters at the t-th iteration, α is the learning rate, and m... t For the corrected first-order moment estimate, v t This is the corrected second-order moment estimate, where ∈ is a small constant used to ensure numerical stability.
8. The torque prediction method for mixing high-viscosity materials as described in claim 2, characterized in that: In step 5.2, the loss value l is determined using the mean square error loss function, as shown in equation (9):
9. The torque prediction method for high-viscosity material mixing as described in claim 2, characterized in that: In step 5.4, the standard mean square error (MSE) is used as the evaluation index for the torque prediction model to measure the average error between the predicted and actual values, as shown in equation (10): Where n is the number of samples, y i This is the actual torque value. The torque value predicted by the torque prediction model; Using the coefficient of determination R 2 The goodness of fit of the regression model is measured as shown in Equation (11): in, The mean of the true values; Coefficient of determination R 2 The closer the value is to 1, the better the torque prediction model fits the data, meaning the higher the proportion of dependent variable variability that the torque prediction model can explain.