Control method for ultrafine grinding link of oolitic hematite

Abstract

The invention provides a control method for superfine grinding processing of oolitic hematite, which is used for controlling grinding equipment in a superfine grinding link, and specifically comprises the following steps: step 1, data acquisition and preprocessing: step 11, acquiring key data of the superfine grinding link in real time through a sensor, the key data comprises particle size distribution, power, current, vibration frequency and amplitude of the ore grinding equipment, and temperature of a bearing and a motor of the ore grinding equipment; 12, cleaning, denoising and integrating the data by using a big data platform; step 13, storing the preprocessed data as training data; 2, constructing a granularity prediction model, an energy consumption optimization model and an equipment health management model, and respectively training the granularity prediction model, the energy consumption optimization model and the equipment health management model by adopting the training data; and step 3, inputting the key data acquired in real time into the granularity prediction model, the energy consumption optimization model and the equipment health management model, controlling the ore grinding equipment according to the output of the granularity prediction model and the energy consumption optimization model, and obtaining an equipment health prediction result according to the output of the equipment health management model. And judging the health state of the equipment according to the equipment health prediction result.

Description

Technical Field

[0001] The present invention relates to the field of oolitic hematite processing, and in particular to a control method for an ultrafine grinding process of oolitic hematite. Background Art

[0002] High-phosphorus oolitic hematite is a complex and difficult iron ore to process. Currently, the main selection method is direct reduction followed by ultrafine grinding followed by magnetic separation. This process allows for the high-value-added utilization of high-phosphorus oolitic hematite, effectively removing harmful phosphorus, and providing alternatives to foreign iron ore for the Chinese steel industry. Ultrafine grinding is a key step in this process, effectively removing harmful phosphorus and improving product quality. Within the ultrafine grinding process, grinding particle size, energy consumption, and equipment health management are key factors affecting process efficiency. Traditional methods rely on manual experience, making precise control and optimization difficult, resulting in high energy consumption, low efficiency, and inconsistent product quality. The development of artificial intelligence (AI) technology has provided new insights for optimizing the ultrafine grinding process. Using techniques such as machine learning and deep learning, intelligent control of the grinding process can be achieved, improving process efficiency and product quality. However, there is no systematic approach to applying AI technology to optimize the ultrafine grinding process of oolitic hematite. Summary of the Invention

[0003] The main purpose of the present invention is to provide a control method for the ultrafine grinding link of oolitic hematite. By introducing an intelligent model, the data monitored in the grinding link is processed and corresponding control parameters are generated to achieve intelligent control of the grinding process.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is: an AI control method for ultrafine grinding of oolitic hematite, which is used to control the grinding equipment in the ultrafine grinding process, and specifically includes the following steps:

[0005] Step 1: Data collection and preprocessing to obtain grinding data, including grinding particle size distribution data, energy consumption data, grinding equipment vibration data, bearing and motor temperature of grinding equipment, grinding time, medium ratio, and equipment failure conditions. Energy consumption data is calculated by collecting real-time voltage and current. Grinding equipment vibration data includes grinding vibration frequency and amplitude. Equipment failure conditions include the number of equipment failures, components of each failure, and maintenance conditions.

[0006] Step 2: Input the obtained grinding data into the control model, and the control model outputs the numerical value of the grinding data to be controlled to the grinding equipment. The control model includes:

[0007] The particle size prediction model takes as input the real-time collected grinding time, media ratio, grinding concentration, particle size distribution and the calculated effective grinding time coefficient, and outputs the grinding time and media ratio;

[0008] Energy consumption optimization model, the input data is the real-time collected grinding concentration, speed, media size, energy consumption and the calculated energy consumption change rate, and the output is grinding concentration and speed;

[0009] The equipment health management model uses real-time collected vibration frequency, temperature, current, and historical equipment failure information as input, and outputs the predicted equipment failure probability and remaining service life.

[0010] Step 3: Input the grinding time, medium ratio, grinding concentration, and rotation speed obtained in step 2 into the grinding equipment to control the grinding process, and maintain the grinding equipment according to the equipment failure probability and the remaining service life of the equipment.

[0011] Preferably, the granularity prediction model adopts a hybrid LSTM-CNN structure. The LSTM is designed to be 3 layers, with 128 neurons in each layer. The time step is set to 10, corresponding to 10 consecutive sampling points. The input data is processed through the following steps:

[0012] The real-time collected grinding time, media ratio, grinding concentration, particle size distribution, and the calculated effective grinding time coefficient are input into the LSTM and CNN respectively. The LSTM outputs the time series characteristics of the grinding parameters, and the CNN uses five 1D convolution kernels to extract the local features of the particle size distribution curve.

[0013] The obtained time series characteristics and local features are spliced ​​together;

[0014] The concatenated features are input into the fully connected layer, and the fully connected layer outputs the grinding time and medium ratio.

[0015] Preferably, the effective grinding time coefficient is calculated by the following formula:

[0016] η=(t actual / t theoretical )×(1-e (-k*ΔD) ),

[0017] Where ΔD is the difference between the current particle size and the target particle size, η is the effective grinding time coefficient, t actual is the actual grinding time, t theoretical is the theoretical grinding time, and k is a constant related to the characteristics of oolitic hematite.

[0018] Preferably, the particle size prediction model also has a loss function, the formula is as follows:

[0019] L = αMSEtime +βMSE ratio +γMSE distribution ,

[0020] Among them, α, β, and γ are weight coefficients optimized for oolitic hematite, and MSE time is the mean square error of grinding time, which is used to measure the difference between the grinding time predicted by the model and the actual grinding time; MSE ratio MSE is the mean square error of the medium ratio, which is used to measure the difference between the medium ratio predicted by the model and the actual medium ratio; distribution is the mean square error of the particle size distribution, which is used to measure the difference between the particle size distribution predicted by the model and the actual particle size distribution.

[0021] Preferably, the energy consumption optimization model includes a random forest model, the input of which is the real-time collected grinding concentration, rotation speed, medium size, energy consumption and the calculated energy consumption change rate, and the output is the preliminary predicted grinding concentration and rotation speed. The energy consumption change rate is calculated using the following formula:

[0022] dE / dx=K×x -n ×f(C,ω),

[0023] Where dE / dx is the energy consumption change rate, x is the particle size, C is the grinding concentration, ω is the rotation speed, and K and n are specific parameters of oolitic hematite.

[0024] Preferably, the energy consumption optimization model also includes a DDQN algorithm model, the input of the DDQN algorithm model is the grinding concentration, rotation speed, energy consumption, productivity, product quality and the preliminary optimized grinding concentration and rotation speed output by the random forest model, and the output is the final optimized grinding concentration and rotation speed, and the grinding equipment is controlled by the final optimized grinding concentration and rotation speed.

[0025] Preferably, the DDQN algorithm model has a reward function for evaluating the grinding process with the final optimized grinding concentration and speed input. The reward function is designed as follows:

[0026] R=λ1(ΔE / E std )+λ2(ΔP / P std )+λ3(QQ min ),

[0027] Where ΔE is the energy consumption reduction, E std is the benchmark energy consumption value, ΔP is the productivity change, P std is the benchmark productivity, Q is the product quality index, Q min It is the minimum threshold of product quality.

[0028] Preferably, the energy consumption optimization model further includes a constraint processing mechanism, which includes a penalty term for processing grinding process constraints, and the formula is as follows:

[0029] Penalty=∑[max(0,g i (x)-b i )] 2 ,

[0030] Among them, g i (x) is the i-th constraint function, the maximum value of i is the number of data in the grinding data that needs to be constrained, x is the i-th data type in the grinding data, b i is the boundary value corresponding to the i-th data type, max(0,g i (x)-b i ) is 0 and g i (x)-b i The maximum value in .

[0031] Preferably, the equipment health management model includes:

[0032] Multimodal fusion model, used to fuse the characteristics of vibration frequency, temperature, current, and equipment failure during the grinding process;

[0033] The fault propagation model uses a Bayesian network to predict the probability of equipment failure through the following steps:

[0034] Step a: Build a Bayesian network:

[0035] Step a1: Analyze the causal relationship between the failures of various components of the equipment;

[0036] Step a2: Based on the fault causal relationship, different fault events and their associated factors are represented in the form of nodes;

[0037] Step a3: Determine the prior probability based on the nodes, quantify the possibility of fault occurrence and its propagation law, achieve early warning of potential faults and accurate location of the root cause of the fault, and build a Bayesian network;

[0038] Step b, training the Bayesian network: using historical data of vibration frequency, temperature, current, and equipment failure as input to train the Bayesian network;

[0039] Step c: After the training is completed, the real-time collected vibration frequency, temperature, current and historical equipment failure conditions are input into the Bayesian network to predict the probability of equipment failure.

[0040] The features fused by the multimodal fusion model are input into the Bayesian network to analyze the causal relationship and propagation path of failures between various equipment components. Different failure events and their associated factors are represented in the form of nodes. The prior probability is determined based on the nodes, and the possibility of failure and its propagation pattern are quantified. This enables early warning of potential failures and accurate location of the root cause of the failure, and outputs the probability of equipment failure.

[0041] The remaining life prediction model uses the LSTM+attention mechanism to input the equipment failure probability predicted by the fault propagation model, historical equipment failure conditions, and real-time detected vibration frequency, temperature, and current into the remaining life prediction model and output the remaining life prediction value RUL t :

[0042] RUL t =∑(a i ×h i )+b,

[0043] Among them, h i is the i-th LSTM state, a i is the weight corresponding to the i-th LSTM state, and b is the bias.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention incorporates AI control into the existing grinding process, inputs real-time data from the grinding process into an algorithm model, and adjusts control data during the grinding process using the data output by the algorithm model, enabling improved grinding efficiency and precision with reduced energy consumption. Special feature engineering utilizes equivalent grinding area calculations for media ratios to ensure that the model-recommended ratios conform to the laws of grinding mechanics.

[0046] (2) Process constraint embedding: Add grinding concentration safety limit to the loss function (C must ∈ [65%, 75%], otherwise the penalty term takes effect).

[0047] (3) Online self-adaptation: The model is fine-tuned every 24 hours using the latest data to adapt to changes in ore hardness (e.g., fluctuations in grinding difficulty due to changes in bedding in oolitic structures). BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0049] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0050] A control method for ultrafine grinding of oolitic hematite is used to control grinding equipment in the ultrafine grinding process, specifically comprising the following steps:

[0051] Step 1: Data collection and preprocessing, specifically including the following steps:

[0052] Step 11, collecting grinding data, the grinding data includes grinding particle size distribution data, energy consumption data, grinding equipment vibration data, grinding equipment bearing and motor temperature, grinding time, medium ratio, equipment failure conditions, energy consumption data is calculated by collecting real-time voltage and current, grinding equipment vibration data includes grinding vibration frequency and amplitude, equipment failure conditions include the number of equipment failures, each failed component and maintenance conditions;

[0053] Step 12: Data preprocessing: Use a big data platform (such as Hadoop or Spark) to clean, denoise, and integrate the grinding data.

[0054] Step 13: Build a data warehouse: Use the pre-processed grinding data as historical data.

[0055] Step 2: Model training and optimization: The model includes a granularity prediction model, an energy consumption optimization model, and an equipment health management model.

[0056] The particle size prediction model uses a hybrid LSTM-CNN architecture. LSTM stands for long short-term memory (LSTM) and CNN stands for convolutional neural network (CNN). The LSTM is designed with three layers, each containing 128 neurons. The time step is set to 10, corresponding to 10 consecutive sampling points. This architecture processes the time series characteristics of the grinding parameters, which are sequences of parameters that continuously change over time during the grinding process. The CNN uses five 1D convolution kernels to extract local features of the particle size distribution curve. These local features can reflect real-time changes in the particle size distribution during the grinding process. The time series characteristics processed by the LSTM and the local features extracted by the CNN are concatenated and input into a fully connected layer, which processes these features to output the grinding time and media ratio.

[0057] The inputs of the particle size prediction model include grinding time, media ratio, grinding concentration, particle size distribution and effective grinding time coefficient, among which the media ratio characteristics are weighted by the formula Σ where w i is the weight proportion of the i-th medium, d iThe diameter is. By adopting the weighted manner, on the one hand, different sizes of media play different roles in the grinding process, and the media with larger diameter mainly crush larger ore particles, and the media with smaller diameter are used to grind smaller particles, and by weighting the weight proportion and diameter of the media, the grinding capacity of the media can be more comprehensively and accurately reflected, so that the model can more accurately predict the grinding effect; on the other hand, the input quality of the model is improved, and the proportions of multiple media are converted into one or several comprehensive values through weighting, which can reduce the dimension of input features, make the model input more concise and efficient, and at the same time retain the key information. The input mode of the model is to calculate a value through a weighting processing formula by using the proportions of several media, and then take the value as an input feature of the model. The value after the weighting processing can comprehensively reflect the matching of the media in the current grinding process, and provide more accurate input information for the model.

[0058] The effective grinding time coefficient is calculated by the following formula:

[0059] η = (t actual / t theoretical ) × (1 - e (-k*ΔD) ), wherein ΔD is the difference between the current particle size and the target particle size, reflecting the progress of the grinding process; η represents the effective grinding time coefficient, used to measure the efficiency ratio of the actual grinding time and the theoretically required time; t actual is the actual grinding time, indicating the actual time spent to complete the grinding operation; t theoretical is the theoretical grinding time, indicating the estimated time required to complete grinding under ideal conditions, and k is a constant related to the characteristics of oolitic hematite. This formula dynamically evaluates the efficiency of the grinding process by combining the actual grinding time and the particle size difference, thereby helping to optimize the grinding parameters, reduce energy consumption and improve product quality. Through this formula, the operator can monitor the grinding efficiency in real time, and adjust the grinding time, media matching and other parameters according to the value of η, to achieve a more optimal grinding effect.

[0060] The output of the particle size prediction model is the grinding time and the media matching, and the output grinding time and the media matching are used to control the grinding process.

[0061] In the training process, whether the training is completed is determined by optimizing the loss function. The loss function adopts mean square error (MSE), and the model parameters are optimized through back propagation. The loss function is L = αMSE time + βMSE ratio + γMSE distribution , wherein α, β and γ are weight coefficients optimized for oolitic hematite, used to balance the contributions of different parts, preferably α = 0.4, β = 0.3 and γ = 0.3, and MSE timeis the mean square error of grinding time, which is used to measure the difference between the grinding time predicted by the model and the actual grinding time; MSE ratio MSE is the mean square error of the medium ratio, which is used to measure the difference between the medium ratio predicted by the model and the actual medium ratio; distribution is the mean squared error of the particle size distribution, which measures the difference between the model's predicted and actual particle size distribution. This multi-objective loss function enables the model to simultaneously optimize grinding time, media ratio, and particle size distribution during training. This integrated optimization approach improves grinding efficiency, reduces energy consumption, and enhances product quality. In practice, the model adjusts its parameters by minimizing this loss function, ensuring that the predicted values ​​are as close as possible to the actual values. In this way, the model learns the complex relationships within the grinding process and provides precise control recommendations in actual production.

[0062] Compared with the general model, this dedicated architecture improves the prediction accuracy by more than 15% in the ultrafine grinding scenario of oolitic hematite.

[0063] For the energy consumption optimization model, the random forest (RF) model and the DDQN algorithm model are used. The input data of the energy consumption optimization model are grinding concentration, rotation speed, medium size, historical data of energy consumption and energy consumption change rate. The output is grinding concentration and rotation speed. The energy consumption change rate is calculated by the grinding kinetic equation. The formula is as follows:

[0064] dE / dx=K×x -n ×f(C,ω),

[0065] Where dE / dx is the energy consumption change rate, x is the particle size, C is the grinding concentration, ω is the rotation speed, K and n are the specific parameters of oolitic hematite, and f(C,ω) is the coupling influence function of grinding concentration and rotation speed. The specific formula is as follows:

[0066]

[0067] C is the current grinding concentration, C opt is the optimal grinding concentration, which can be determined by experiments or by past experience. crit is the critical speed, which is the limit value to avoid the centrifugal movement of the grinding equipment. a, b, and c are dynamic fitting parameters obtained through random forest model training.

[0068] The random forest model is trained by inputting historical data of grinding concentration, speed, medium size, energy consumption and energy consumption change rate into the random forest model. The random forest model learns the relationship between grinding parameters and energy consumption, and outputs the predicted energy consumption value, the preliminary optimized grinding concentration and speed.

[0069] The DDQN algorithm model is trained. The input data during training are: grinding concentration, speed, energy consumption, productivity, product quality, and the preliminary optimized grinding concentration and speed output by the random forest model. The output data are the final optimized grinding concentration and speed, which are used to control the grinding process.

[0070] The DDQN algorithm model has a reward function, which is used to evaluate the grinding process of the final optimized grinding concentration and speed input. The reward function is designed as follows:

[0071] R=λ1(ΔE / E std )+λ2(ΔP / P std )+λ3(QQ min ),

[0072] Where ΔE is the energy consumption reduction, E std is the benchmark energy consumption value, ΔP is the productivity change, P std is the benchmark productivity, Q is the product quality index, Q min represents the minimum threshold for product quality. During energy optimization, the model monitors and evaluates the energy efficiency of the current grinding operation in real time. For example, as the energy reduction ΔE increases, it indicates that the optimization measures implemented by the model have effectively reduced energy consumption. Accordingly, the value of the reward function R also increases, encouraging the model to continue using this strategy. Based on the principle of maximizing cumulative rewards, the model continuously adjusts its strategy to achieve energy optimization while balancing productivity and product quality.

[0073] During the training process of the energy consumption optimization model, a loss function is introduced. The loss function formula is as follows:

[0074]

[0075] Where m is the number of training data groups, is the energy consumption value predicted by the random forest model when the i-th group of training data is input, is the energy consumption value calculated by monitoring the current and voltage, and Penalty is the penalty term. The penalty term formula is as follows:

[0076] Penalty=∑[max(0,g i (x)-b i )] 2 ,

[0077] Among them, g i (x) is the i-th constraint function. The specific calculation method of the constraint function can adopt the existing technology. The maximum value of i is the number of data in the grinding data that needs to be constrained. x is the i-th data type in the grinding data. b iis the boundary value corresponding to the i-th data type, max(0,g i (x)-b i ) is 0 and g i (x)-b i The maximum value in . During model training, if a constraint is violated (e.g., the rotational speed exceeds the safe range), the penalty term increases the loss function, and the model adjusts the parameters through backpropagation. In actual grinding, the model adjusts the grinding operation based on the optimized parameters, ensuring energy consumption optimization while meeting process constraints and ensuring a safe and stable grinding process.

[0078] For the equipment health management model, it includes:

[0079] Multimodal fusion model, used to fuse the characteristics of vibration frequency, temperature, current, and equipment failure during the grinding process;

[0080] The fault propagation model uses a Bayesian network for prediction. Before using the Bayesian network, you need to build the Bayesian network:

[0081] Step 1: Analyze the causal relationship between the failures of various components of the equipment;

[0082] Step 2: Based on the fault causal relationship, different fault events and their associated factors are represented in the form of nodes;

[0083] Step 3: Determine the prior probability based on the node, quantify the possibility of fault occurrence and its propagation law, and achieve early warning of potential faults and accurate location of the root cause of the fault.

[0084] The constructed Bayesian network outputs the device failure probability, which includes the failure probability of each component in the device that needs to be predicted.

[0085] After constructing the Bayesian network, historical data of vibration frequency, temperature, current, and equipment failure conditions are used as input to train the Bayesian network.

[0086] After the training is completed, the probability of equipment failure can be predicted by inputting the real-time collected vibration frequency, temperature, current and historical equipment failure conditions.

[0087] The remaining life prediction model uses the LSTM+attention mechanism to input the output of the fault propagation model and historical data into the remaining life prediction model to predict the remaining service life of the equipment. The attention mechanism is used to assign weights to different LSTM hidden states, highlighting the impact of key features on life prediction. The remaining life prediction value RUL of the grinding equipment is output through the following formula t =∑(a i ×h i )+b,h iFor the i-th LSTM state, a i For the i-th LSTM state, a

[0088] The device health management model generates device health state evaluation, failure prediction, remaining life prediction, and maintenance recommendations based on the device failure probability obtained from the failure propagation model and the remaining life prediction model.

[0089] The working steps of the health management model are as follows:

[0090] Step 1: The multi-modal fusion model fuses the vibration frequency, temperature, current, and device failure conditions in the grinding process.

[0091] Step 2: The fused features are input into the failure propagation model, which predicts the failure probability based on the input fused features to obtain the device failure prediction probability.

[0092] Step 3: The predicted failure probability and device failure conditions are input into the remaining life prediction model, which predicts the remaining life of the grinding device to obtain the predicted remaining life of the grinding device.

[0093] Step 4: Based on the device failure prediction probability and the predicted remaining life of the grinding device, the device health state evaluation and maintenance recommendations are obtained, and the device failure prediction probability, the predicted remaining life of the grinding device, the device health state evaluation, and the maintenance recommendations are output together.

[0094] Step 3: Real-time control and decision-making:

[0095] The real-time collected grinding time, medium ratio, grinding concentration, particle size distribution, and calculated effective grinding time coefficient are input into the particle size prediction model, which outputs the required grinding time and medium ratio. The grinding device is controlled according to the grinding time and medium ratio output by the particle size prediction model.

[0096] The real-time collected grinding concentration, rotational speed, medium size, energy consumption, and calculated energy consumption change rate are input into the energy consumption optimization model, which outputs the grinding concentration and rotational speed. The grinding device is controlled according to the grinding concentration and rotational speed output by the energy consumption optimization model.

[0097] The vibration frequency, temperature, current, and device failure conditions are input into the device health management model, which predicts the device health prediction results. The grinding device is maintained according to the device health prediction results. Maintenance includes maintenance and repair.

[0098] Step 4: Process effect evaluation and feedback:

[0099] Step 41: Obtain particle size qualification rate, energy efficiency, and equipment failure rate;

[0100] Step 42: determine whether the particle size qualification rate, energy efficiency and equipment failure rate meet the requirements. If not, continue to train the corresponding model until the particle size qualification rate, energy efficiency and equipment failure rate meet the requirements.

[0101] Experimental example

[0102] This experimental example shows the specific application of the particle size prediction model in the ultrafine grinding of oolitic hematite.

[0103] 1. Implementation Background: During the ultrafine grinding of oolitic hematite, grinding particle size directly impacts the dephosphorization effectiveness and iron concentrate grade in the subsequent magnetic separation process. Traditional manual control methods make it difficult to adjust grinding parameters in real time, resulting in large particle size fluctuations and high energy consumption. This example utilizes an LSTM-CNN hybrid model, combined with the grinding characteristics of oolitic hematite, to accurately predict and optimize particle size distribution.

[0104] 2. Model construction and training:

[0105] (1) Data preparation:

[0106] Input data (sampling frequency: 1 time / minute): Grinding time (t): current grinding time (min); Media ratio (w1:w2:w3): weight ratio of steel balls (Φ10mm), steel segments (Φ15mm), and ceramic balls (Φ8mm); Grinding concentration (C): slurry solid content (%); Particle size distribution (D50, D90): real-time detection by laser particle size analyzer.

[0107] Output data: recommended grinding time (Δt) at the next moment; optimized media ratio (w1':w2':w3').

[0108] (2) Model architecture:

[0109] The LSTM-CNN hybrid network is used with the following structure:

[0110] (a) LSTM layer (time series feature extraction): 3 layers of LSTM, 128 neurons per layer, Dropout = 0.2; Input: Grinding parameter sequence of the past 10 minutes (sliding window).

[0111] (b) CNN layer (particle size distribution feature extraction): 1D convolution kernel (kernel_size = 3, filters = 32), MaxPooling (pool_size = 2); Input: particle size distribution curve (D 50 、D 90 etc. 10 feature points).

[0112] (c) Fully connected layer (decision output): outputs the grinding time adjustment (Δt) and the optimized media ratio.

[0113] (3) Loss function:

[0114] The weighted mean square error (WMSE) was used to adjust the weights for the oolitic hematite characteristics:

[0115] L = 0.5·MSE(Δt) + 0.3·MSE(w / ) + 0.2·MSE(D90), where the error weight of D90 is higher to ensure that the coarse particles are fully ground.

[0116] 3. Real-time control process:

[0117] (1) Data acquisition: The sensor collects mill current, vibration, and temperature data in real time, and the particle size analyzer updates D50 and D90 every 30 seconds.

[0118] (2) Model prediction: Input the current data into the trained LSTM-CNN model and output Δt and w'; Example: Current medium ratio: 50% steel balls, 30% steel segments, 20% ceramic balls; Model output: Adjust to 45% steel balls, 35% steel segments, 20% ceramic balls (reduce the proportion of steel balls to reduce over-wear).

[0119] (3) Equipment control: PLC automatically adjusts the mill feed rate and media replenishment to ensure that the particle size D90 ≤ 25 μm (dephosphorization requirement).

[0120] 4. Implementation effect:

[0121] Comparative test in an oolitic hematite concentrator (traditional manual control vs. AI control):

[0122] Production Index Comparison Table index Traditional methods AI control Improvement effect Particle size pass rate (D 90 ≤ 25 μm) 78％ 95％ +17％ Energy consumption per ton of ore (kWh / t) 42.5 38.2 -10％ Grinding mill failure rate (times / month) 3.2 1.1 -66％

[0123] 5. Conclusion

[0124] This embodiment uses the LSTM-CNN hybrid model + oolitic hematite special optimization strategy to achieve: precise particle size control (D 90 The fluctuation range has been reduced from ±5μm to ±1.5μm); energy consumption has been reduced by 10% (by dynamically adjusting the media ratio to reduce ineffective grinding); and equipment life has been extended (health model-based early warning has extended the bearing replacement cycle from 6 months to 9 months).

[0125] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A control method for ultrafine grinding of oolitic hematite, for controlling a grinding device in an ultrafine grinding process, specifically comprising the following steps: Step 1: Data collection and preprocessing to obtain grinding data, including grinding particle size distribution data, energy consumption data, grinding equipment vibration data, bearing and motor temperature of grinding equipment, grinding time, medium ratio, and equipment failure conditions. Energy consumption data is calculated by collecting real-time voltage and current. Grinding equipment vibration data includes grinding vibration frequency and amplitude. Equipment failure conditions include the number of equipment failures, components of each failure, and maintenance conditions. Step 2: Input the obtained grinding data into the control model, and the control model outputs the numerical value of the grinding data to be controlled to the grinding equipment. The control model includes: The particle size prediction model takes as input the real-time collected grinding time, media ratio, grinding concentration, particle size distribution and the calculated effective grinding time coefficient, and outputs the grinding time and media ratio; Energy consumption optimization model, the input data is the real-time collected grinding concentration, speed, media size, energy consumption and the calculated energy consumption change rate, and the output is grinding concentration and speed; The equipment health management model uses real-time collected vibration frequency, temperature, current, and historical equipment failure information as input, and outputs the predicted equipment failure probability and remaining service life. Step 3: Input the grinding time, medium ratio, grinding concentration, and rotation speed obtained in step 2 into the grinding equipment to control the grinding process, and maintain the grinding equipment according to the equipment failure probability and the remaining service life of the equipment.

2. The control method for ultrafine grinding of oolitic hematite according to claim 1, characterized in that: The particle size prediction model adopts a hybrid LSTM-CNN structure. The LSTM is designed to be three-layer, with 128 neurons in each layer. The time step is set to 10, corresponding to 10 consecutive sampling points, which is used to process the time series characteristics of grinding parameters. The CNN uses five 1D convolution kernels to extract the local features of the particle size distribution curve.

3. The control method for ultrafine grinding of oolitic hematite according to claim 1, characterized in that: The effective grinding time coefficient is calculated by the following formula: η=(t actual / t theoretical )×(1-e (-k*ΔD) ), Where ΔD is the difference between the current particle size and the target particle size, η is the effective grinding time coefficient, t actual is the actual grinding time, t theoretical is the theoretical grinding time, and k is a constant related to the characteristics of oolitic hematite.

4. The AI ​​control method for ultrafine grinding of oolitic hematite according to claim 1, characterized in that: The particle size prediction model also has a loss function, which is as follows: L=αMSE time +βMSE ratio +γMSE distribution , Among them, α, β, and γ are weight coefficients optimized for oolitic hematite, and MSE time is the mean square error of grinding time, which is used to measure the difference between the grinding time predicted by the model and the actual grinding time; MSE ratio MSE is the mean square error of the medium ratio, which is used to measure the difference between the medium ratio predicted by the model and the actual medium ratio; distribution is the mean square error of the particle size distribution, which is used to measure the difference between the particle size distribution predicted by the model and the actual particle size distribution.

5. The AI ​​control method for ultrafine grinding of oolitic hematite according to claim 1, characterized in that: The energy consumption optimization model includes a random forest model. The input of the random forest model is the real-time collected grinding concentration, rotation speed, medium size, energy consumption and the calculated energy consumption change rate. The output is the preliminary predicted grinding concentration and rotation speed. The energy consumption change rate is calculated using the following formula: dE / dx=K×x -n ×f(C,ω), Where dE / dx is the energy consumption change rate, x is the particle size, C is the grinding concentration, ω is the rotation speed, and K and n are specific parameters of oolitic hematite.

6. A control method for ultrafine grinding of oolitic hematite according to claim 5, characterized in that: The energy consumption optimization model also includes a DDQN algorithm model, the input of which is grinding concentration, rotation speed, energy consumption, productivity, product quality, and the preliminary optimized grinding concentration and rotation speed output by the random forest model, and the output is the final optimized grinding concentration and rotation speed. The grinding equipment is controlled by the final optimized grinding concentration and rotation speed.

7. A control method for ultrafine grinding of oolitic hematite according to claim 6, characterized in that: The DDQN algorithm model has a reward function, which is used to evaluate the grinding process of the final optimized grinding concentration and speed input. The reward function is designed as follows: R=λ1(ΔE / E std )+λ2(ΔP / P std )+λ3(QQ min ), Where ΔE is the energy consumption reduction, E std is the benchmark energy consumption value, ΔP is the productivity change, P std is the benchmark productivity, Q is the product quality index, Q min It is the minimum threshold of product quality.

8. The control method for ultrafine grinding of oolitic hematite according to claim 6, characterized in that: The energy consumption optimization model also has a constraint processing mechanism, which has a penalty term for processing grinding process constraints. The formula is as follows: Penalty=∑[max(0,g i (x)-b i )] 2 , Among them, g i (x) is the i-th constraint function, the maximum value of i is the number of data in the grinding data that needs to be constrained, x is the i-th data type in the grinding data, b i is the boundary value corresponding to the i-th data type, max(0,g i (x)-b i ) is 0 and g i (x)-b i The maximum value in .

9. The control method for ultrafine grinding of oolitic hematite according to claim 1, characterized in that: The equipment health management model includes: Multimodal fusion model, used to fuse the characteristics of vibration frequency, temperature, current, and equipment failure during the grinding process; The fault propagation model uses a Bayesian network to input the features fused by the multimodal fusion model into the Bayesian network to predict the probability of equipment failure; The remaining life prediction model uses the LSTM+attention mechanism to input the equipment failure probability predicted by the fault propagation model, historical equipment failure conditions, and real-time detected vibration frequency, temperature, and current into the remaining life prediction model and output the remaining life prediction value RUL t : NUMBER t =∑(a i ×h i )+b, Among them, h i is the i-th LSTM state, a i is the weight corresponding to the i-th LSTM state, and b is the bias.

10. The control method for ultrafine grinding of oolitic hematite according to claim 9, characterized in that: The fault propagation model predicts the equipment failure rate through the following steps: Step a: Build a Bayesian network: Step a1: Analyze the causal relationship between the failures of various components of the equipment; Step a2: Based on the fault causal relationship, different fault events and their associated factors are represented in the form of nodes; Step a3: Determine the prior probability based on the nodes, quantify the possibility of fault occurrence and its propagation law, achieve early warning of potential faults and accurate location of the root cause of the fault, and build a Bayesian network; Step b, training the Bayesian network: using historical data of vibration frequency, temperature, current, and equipment failure as input to train the Bayesian network; Step c: After the training is completed, the real-time collected vibration frequency, temperature, current and historical equipment failure conditions are input into the Bayesian network to predict the probability of equipment failure.

Images