Multivariable self-optimization control coal mill intelligent agent operation method

Abstract

The invention belongs to the field of artificial intelligence, particularly relates to a multivariable self-optimization control coal mill agent operation method, and aims to solve the problems of variable coupling interference, parameter solidification lag and poor working condition adaptability in traditional control. According to the strategy, an attention mechanism-based LSTM dynamic model is constructed, and future states of six types of variables such as air volume and temperature are predicted; defining a weighted optimization function fusing powder making efficiency, energy consumption, wear and stability, and introducing a soft constraint to guarantee a safety boundary; an optimal control combination is searched on line every five seconds by adopting a depth deterministic strategy gradient algorithm, and the four-dimensional action space of an air door, a hot blast valve, a coal feeder and a grinding roller is covered; and triggering an incremental learning fine tuning model through residual feedback to prevent disastrous forgetting. According to the scheme, multivariable cooperation, online self-optimization and intelligent safety control under all working conditions are realized, and the dynamic response capability and the comprehensive operation efficiency of the coal mill are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method for operating a coal mill intelligent agent with multivariable self-optimization control. Background Technology

[0002] With the accelerated advancement of industrial intelligence, the operation and control of coal mills in coal-fired power generation systems is gradually shifting from being dominated by human experience to a data-driven intelligent decision-making model. Traditional coal mill control systems rely heavily on fixed parameter settings and single-variable feedback regulation, making it difficult to cope with multiple dynamic disturbances such as coal quality fluctuations, load changes, and equipment aging. This results in unstable pulverizing efficiency, high energy consumption, and accelerated equipment wear. Especially under high load or variable operating conditions, the system needs to simultaneously coordinate multiple key variables such as air volume, temperature, current, and vibration, which places higher demands on the adaptability and multi-objective collaborative optimization capabilities of the control strategy.

[0003] However, existing control models generally lack the ability to dynamically model multivariable coupling relationships, leading to lags or oscillations in the regulation process and affecting system stability. Simultaneously, the implicit nonlinear characteristics and operating condition transition patterns in the operating data are not effectively mined, making it difficult for traditional methods to achieve online self-optimization of control parameters, causing the system to operate in a suboptimal state for extended periods. Furthermore, control strategy updates rely on offline simulation or manual tuning, failing to respond in real-time to equipment state evolution and external disturbances, resulting in the continuous accumulation of energy efficiency losses and safety risks.

[0004] Therefore, a multivariable self-optimizing control method for operating a coal mill intelligent agent is desired. Summary of the Invention

[0005] The purpose of this invention is to provide a multivariable self-optimizing control method for operating a coal mill intelligent agent, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multivariable self-optimizing control method for operating a coal mill intelligent agent includes the following specific steps: Step (1) Constructing a multivariable coupled dynamic model: Based on the historical operating data of the coal mill, time series data of six key variables, including air volume, temperature, current, vibration, coal powder fineness, and inlet negative pressure, are collected. The data are preprocessed using a time series segmentation method with a sliding window length of 128, and a nonlinear dynamic coupling relationship model between variables is established through a long short-term memory neural network based on an attention mechanism. The predicted values ​​of each variable in the next control cycle are output; Step (2) Determining Multi-objective optimization function: Based on four core indicators—pulverizing efficiency, unit energy consumption, equipment wear rate, and system stability—a weighted multi-objective optimization function is constructed. The weight coefficients for pulverizing efficiency, unit energy consumption, equipment wear rate, and system stability are 0.4, 0.3, 0.2, and 0.1, respectively. Soft constraints are introduced to limit the adjustment range of air volume and temperature to ensure that the control commands are within the safe operating range of the equipment. Step (3) Execute online strategy optimization: The predicted value output in step (1) is used as the state input, combined with the current operating conditions. The optimal combination of control parameters is searched in the action space using a deep deterministic strategy gradient algorithm. The action space has a dimension of 4, corresponding to the opening of the primary damper, the opening of the hot air regulating valve, the speed of the coal feeder, and the loading force of the grinding roller. The optimization process is executed once every 5 seconds, and each iteration does not exceed 200 milliseconds. Step (4) Implement closed-loop feedback correction: The optimized control parameters are sent to the actuator, and the actual operation feedback data is collected synchronously. The residual vector between the predicted value and the measured value is calculated. If the absolute value of the residual vector exceeds the set threshold of 0.05, the online fine-tuning mechanism of the model is triggered. The weight parameters of the long short-term memory neural network are updated using the incremental learning method. The fine-tuning process uses the most recent 2000 sets of data in the sliding window as training samples. Step (5) Achieve adaptive migration of working conditions: Establish a library of operating modes including low load, medium load, high load and variable load. By calculating the distance between the current operating state and the center of each mode library in real time, the nearest neighbor classification criterion is used to identify the current working condition category, and the corresponding optimization function weight coefficient and action space boundary are dynamically switched to ensure that the control strategy maintains optimal performance in different load ranges.

[0007] Preferably, in step (1), the historical running data sampling frequency is once per second, the data cleaning process removes time periods with a missing rate greater than 5%, and outliers are corrected using the three-standard-deviation method to ensure the integrity and reliability of the input data.

[0008] Preferably, in step (1), the long short-term memory neural network based on the attention mechanism contains 3 hidden layers, each layer has 256 neurons, 8 attention heads, an input sequence length of 64, an output prediction step size of 10, and the model training uses a weighted loss function of mean squared error and mean absolute error with a weight ratio of 7 to 3.

[0009] Preferably, in step (2), the pulverizing efficiency is calculated by the ratio of qualified coal powder output to raw coal consumption per unit time, the unit energy consumption is the ratio of total power consumption of the coal mill to coal powder output, the equipment wear rate is evaluated by the weighted index of the effective value of vibration and the standard deviation of current fluctuation, and the system stability is measured by the reciprocal of the standard deviation of wind temperature and wind pressure.

[0010] Preferably, in step (3), the exploration noise of the deep deterministic policy gradient algorithm is generated by the Ornstein-Uhlenbeck process, the noise attenuation coefficient is 0.995, the target network update cycle is 100 steps, and the experience replay buffer capacity is 10,000 sets of state-action-reward-next state quadruples.

[0011] Preferably, in step (3), the action space boundary is set according to the equipment nameplate parameters and historical safe operation records. The primary air damper opening range is 30% to 90%, the hot air regulating valve opening range is 20% to 85%, the coal feeder speed range is 300 rpm to 900 rpm, and the grinding roller loading force range is 5 kN to 15 kN.

[0012] Preferably, the threshold values ​​for the residual vector in step (4) are configured for different variables respectively: the residual threshold value for air volume is 0.04, the residual threshold value for temperature is 0.06, the residual threshold value for current is 0.05, the residual threshold value for vibration is 0.07, the residual threshold value for coal powder fineness is 0.08, and the residual threshold value for inlet negative pressure is 0.05.

[0013] Preferably, in step (4), the incremental learning adopts the elastic weight solidification regularization method to retain the network connection weights that are important for historical conditions, and to prevent the model from catastrophic forgetting during the learning process of new data. The regularization coefficient is set to 0.1.

[0014] Preferably, in step (5), the operating mode library performs unsupervised division of historical operating condition data using the K-means clustering algorithm. The clustering features include load rate, coal grindability index, ambient temperature and coal mill outlet temperature. The number of clusters is fixed at 4, and the mode center is represented by the mean vector of the samples within each cluster.

[0015] Preferably, the working condition identification process in step (5) is executed once every 10 seconds. If the identification results are consistent for three consecutive times, the working condition switch is confirmed to avoid frequent mode jumps caused by instantaneous disturbances and to ensure the continuity and robustness of the control strategy.

[0016] Preferably, it also includes: establishing a coal mill health status assessment module, calculating the equipment degradation index based on vibration spectrum characteristics and current harmonic components, and automatically adjusting the weight coefficient of equipment wear rate in the optimization function to 0.3 when the degradation index exceeds 0.75, so as to prioritize the safe operation of the equipment.

[0017] Preferably, it also includes: configuring a human-machine collaborative intervention interface, which generates a strategy deviation alarm when the system detects that the optimization results of the control parameters deviate from the operator's historical operating habits by more than 20% for 5 consecutive times, and provides parameter adjustment suggestions for operators to confirm or override.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The ability to coordinate multivariate control has been significantly improved. By constructing a dynamic model based on a long short-term memory neural network with an attention mechanism, the nonlinear coupling relationships among six types of variables, including air volume, temperature, current, and vibration, are accurately characterized, solving the problems of regulation lag and system oscillation caused by mutual interference between variables in traditional single-variable feedback control. Online strategy optimization is executed every 5 seconds, covering the action space of four key actuators, achieving multi-objective, multi-degree-of-freedom collaborative control, enabling the pulverizing system to maintain stable and efficient operation under varying operating conditions.

[0019] Control parameters achieve true online self-optimization. Abandoning the traditional approach of relying on offline simulation or manual tuning, this invention employs a closed-loop feedback correction and incremental learning mechanism, enabling the control model to respond in real time to dynamic disturbances such as coal quality fluctuations and equipment aging. When the prediction residual exceeds a threshold, the system automatically triggers model fine-tuning, updating network weights using the most recent 2000 sets of operational data to ensure that the control strategy always matches the current equipment state, preventing the system from operating under suboptimal conditions for extended periods.

[0020] Operating condition adaptive capability covers the entire load operating range By establishing a library of operating modes encompassing four typical operating conditions and combining this with the nearest neighbor classification criterion to achieve automatic operating condition identification and strategy migration, this invention can dynamically adjust the weights and action boundaries of the optimization function. It prioritizes energy consumption optimization under low load conditions, prioritizes pulverizing efficiency and system stability under high load conditions, and smoothly switches control logic during load changes, significantly improving the overall performance of the coal mill across the entire operating range.

[0021] Dual protection of security and human-machine collaboration mechanism An equipment degradation index assessment module is introduced, which automatically increases the wear rate weight when the equipment condition deteriorates, prioritizing the protection of critical components. At the same time, a human-machine collaborative intervention interface is set up to alarm and confirm optimization results that deviate significantly from operating habits. This not only leverages the autonomous decision-making advantages of the intelligent agent, but also retains the ultimate control of the operators, effectively balancing the relationship between automation and safety. Attached Figure Description

[0022] Figure 1 This is a flowchart of the overall technical solution of the present invention. Detailed Implementation

[0023] Example 1 Please refer to Figure 1 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0024] Currently, coal mill operation suffers from problems such as strong multivariate coupling, frequent changes in operating conditions, and difficulty in real-time optimization of control parameters, leading to large fluctuations in pulverizing efficiency, high energy consumption, accelerated equipment wear, and insufficient system stability. To address these technical problems, this invention proposes a multivariate self-optimizing control method for operating a coal mill intelligent agent, and applies it to this method.

[0025] In the above-mentioned multivariable self-optimizing control method for coal mill intelligent operation, step (1) involves constructing a multivariable coupled dynamic model: based on historical coal mill operation data, time series data of six key variables, including air volume, temperature, current, vibration, coal powder fineness, and inlet negative pressure, are collected. The data is preprocessed using a time series segmentation method with a sliding window length of 128, and a nonlinear dynamic coupling relationship model between variables is established through a long short-term memory neural network based on an attention mechanism. The predicted values ​​of each variable in the next control cycle are then output. Specifically, in step (1), the historical operation data is sampled once per second. During the data cleaning process, time periods with a missing rate greater than 5% are removed, and outliers are corrected using the three-standard-deviation method to ensure the integrity and reliability of the input data. In step (1), the long short-term memory neural network based on an attention mechanism contains three hidden layers, each with 256 neurons, 8 attention heads, an input sequence length of 64, and an output prediction step size of 10. The model training uses a weighted loss function of mean square error and mean absolute error, with a weight ratio of 7:3. The remote sensing image data in step (1) includes high-resolution optical images, synthetic aperture radar images, and hyperspectral images, with a temporal resolution of no less than 15 days and a spatial resolution better than 2 meters. Multi-temporal image alignment is achieved through geometric correction and radiometric normalization processing based on control points. The real estate registration data in step (1) covers property rights information of natural resources such as land, forest land, water areas, and minerals. Entities are associated through unified social credit codes and spatial location codes to ensure accurate matching of property rights boundaries and geographic coordinates, with a matching error of less than 0.5 meters. In step (1), the data acquisition module consists of an industrial-grade sensor array, including an air volume sensor (model: ABB 800XA), a temperature sensor (model: PT100), a current transformer (model: LZZBJ9-10), a vibration accelerometer (model: PCB 356A15), a coal powder fineness online analyzer (model: MettlerToledo FMS-200), and an inlet negative pressure transmitter (model: Endress+Hauser DMP336). All sensors are connected to the distributed control system (DCS) and synchronized to the central database. The data transmission protocol uses Modbus TCP / IP, the communication rate is set to 10 Mbps, the data packet size is 1024 bytes, a complete state vector is generated every second, and the storage format is a CSV file. The field structure includes timestamp (accurate to milliseconds), variable name, value, unit, and quality flag (0 indicates valid, 1 indicates invalid). In step (1), the time series segmentation method with a sliding window length of 128 divides the original time series into continuous and overlapping subsequences. Each subsequence contains 128 sampling points, and the sliding step between adjacent windows is 1, forming a set of time series segments for subsequent model training and inference.In step (1), the data cleaning process first checks the integrity of the original data. If the missing rate of a certain variable exceeds 5% in 128 consecutive sampling points, the data in that time period is marked as invalid and removed from the training set. For outliers in a single sampling point, the three-standard-deviation method is used for identification and correction. That is, the mean μ and standard deviation σ of the variable in the sliding window are calculated. If the value x of a certain point satisfies |x - μ|>3σ, it is replaced with the boundary value in μ + 3σ or μ - 3σ to eliminate extreme noise interference. The long short-term memory neural network structure based on the attention mechanism in step (1) consists of an input layer, three stacked LSTM layers, an attention mechanism layer, a fully connected layer, and an output layer. The input layer receives a time sequence of length 64, with each time step containing a 6-dimensional feature vector corresponding to air volume, temperature, current, vibration, coal powder fineness, and inlet negative pressure. The input sequence is mapped to a 128-dimensional dense vector space by the embedding layer and then fed into the first LSTM layer. Each LSTM layer contains 256 memory units, using the tanh activation function and sigmoid gating mechanism to hide the state. The state and cell state are transmitted between layers. The sequence output by the third LSTM layer is used as the query and key input for the attention mechanism. There are 8 attention heads, and each head independently calculates the attention weight. After normalization by the softmax function, the weighted sum is used to obtain the context vector. This vector is then mapped to the output layer through two fully connected network layers (128 neurons per layer, ReLU activation). The output layer is a 6-dimensional vector, which corresponds to the predicted values ​​of each variable in the next control cycle. The prediction step size is 10, that is, the model can predict the state of the next 10 time steps at the same time. In step (1), the model training adopts a weighted loss function of mean squared error (MSE) and mean absolute error (MAE) with a weight ratio of 7:3. The expression of the loss function is:

[0026] in, The calculation formula is:

[0027] The formula for calculating MAE is:

[0028] For the sample size, For the true value, These are predicted values. The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. An early stopping mechanism was implemented, terminating training if the loss on the validation set did not decrease after 10 consecutive epochs to prevent overfitting. After training, the model parameters were saved in TensorFlow SavedModel format and deployed on edge computing nodes to support low-latency inference services.

[0029] In the above-mentioned multivariate self-optimization control method for operating a coal mill intelligent agent, step (2) defines a multi-objective optimization function: based on four core indicators—pulverizing efficiency, unit energy consumption, equipment wear rate, and system stability—a weighted multi-objective optimization function is constructed, where the weight coefficient for pulverizing efficiency is 0.4, the weight coefficient for unit energy consumption is 0.3, the weight coefficient for equipment wear rate is 0.2, and the weight coefficient for system stability is 0.1. Soft constraint boundaries are introduced to limit the adjustment range of air volume and temperature, ensuring that the control commands are within the safe operating range of the equipment. Specifically, in step (2), pulverizing efficiency is calculated as the ratio of qualified coal powder production to raw coal consumption per unit time; unit energy consumption is the ratio of total power consumption of the coal mill to coal powder production; equipment wear rate is evaluated based on the weighted index of the effective value of vibration and the standard deviation of current fluctuation; and system stability is measured by the reciprocal of the standard deviation of air temperature and air pressure. In step (2), the graph neural network employs a graph attention mechanism. The node feature vector dimension is 128, and the edge weights are calculated by multiplying the spatial distance decay function with the functional similarity. The decay coefficient is set to 0.8, and the similarity threshold is set to 0.7. In step (2), the spatial autocorrelation analysis uses a combination of the Moran index and the Gillley index for discrimination. The calculation window radius is set to 5 kilometers, and the significance level is set to 0.05. This is used to identify high-high and low-low clustering regions as the initial connection basis for the association graph. The multi-objective optimization function in step (2) is as follows:

[0030] in, To improve powder production efficiency, Energy consumption per unit For equipment wear rate, For the maximum permissible wear rate, For wind temperature standard deviation, For wind pressure standard deviation. Pulverizing efficiency. The calculation formula is:

[0031] in, Qualified coal powder output (unit: kg / h). Density of pulverized coal (unit: kg / m³). The qualified rate of pulverized coal (unit: %) This refers to raw coal consumption (unit: kg / h). Unit energy consumption. The calculation formula is:

[0032] in, Total power consumption of the coal mill (unit: kWh / h). Coal powder production (unit: kg / h). Equipment wear rate. The calculation formula is:

[0033] in, The effective value of vibration (unit: mm / s). Reference vibration value (unit: mm / s). The standard deviation of current fluctuation (unit: A) The reference current fluctuation standard deviation (unit: A) This is a weighting factor, with a value of 0.6. System stability is measured by the reciprocal of the standard deviations of wind temperature and wind pressure, i.e.:

[0034] in, The standard deviation of air temperature (unit: °C) The standard deviation of wind pressure is expressed in Pa. In step (2), the soft constraint boundary limits the adjustment range of air volume and temperature. The air volume adjustment range is set to 30% to 90%, and the temperature adjustment range is set to 180°C to 320°C. Control commands exceeding this range will be truncated to the boundary value to prevent the equipment from operating beyond its limits. The weight coefficients of the optimization function in step (2) can be dynamically adjusted according to the operating conditions. For example, in the low load stage, the weight of unit energy consumption can be increased to 0.4, and the weight of pulverizing efficiency can be reduced to 0.3 to prioritize reducing energy consumption; in the high load stage, the weight of pulverizing efficiency can be increased to 0.5, and the weight of system stability can be increased to 0.15 to ensure output and stability.

[0035] In the above-mentioned multivariable self-optimizing control method for operating a coal mill intelligent agent, step (3) involves performing online strategy optimization: the predicted value output in step (1) is used as the state input, combined with the current working condition label, and a deep deterministic strategy gradient algorithm is used to search for the optimal combination of control parameters in the action space. The action space has a dimension of 4, corresponding to the primary damper opening, hot air regulating valve opening, coal feeder speed, and grinding roller loading force, respectively. The optimization process is performed once every 5 seconds, and each iteration does not exceed 200 milliseconds. Specifically, the exploration noise of the deep deterministic strategy gradient algorithm in step (3) is generated using the Ornstein-Uhlenbeck process, with a noise attenuation coefficient of 0.995. The target network update cycle is 100 steps, and the experience replay buffer capacity is 10,000 sets of state-action-reward-next state quadruplets. In step (3), the action space boundary is set according to the equipment nameplate parameters and historical safe operation records. The primary air damper opening range is 30% to 90%, the hot air regulating valve opening range is 20% to 85%, the coal feeder speed range is 300 rpm to 900 rpm, and the grinding roller loading force range is 5 kN to 15 kN. In step (3), the ecological security target uses the ecological carrying capacity index as a quantitative indicator, the economic efficiency target uses the unit resource output ratio as a quantitative indicator, and the social equity target uses the inverse of the Gini coefficient as a quantitative indicator. The three constitute a three-dimensional optimization target space. In step (3), the improved multi-objective particle swarm optimization algorithm introduces an elite archiving mechanism and adaptive inertia weight. The population size is set to 200, the maximum number of iterations is set to 500, and the convergence criterion is that the Pareto front change rate is less than 1% for 50 consecutive generations. In step (3), the state space of the deep deterministic policy gradient algorithm is formed by concatenating the 6-dimensional predicted value vector output from step (1) with the current working condition label (category encoding), resulting in a 10-dimensional state vector. This vector is then input into the policy network, which is a three-layer fully connected network with 128 neurons per layer, using the ReLU activation function. The output layer is a 4-dimensional action vector corresponding to four control parameters. In step (3), the exploration noise is generated using the Ornstein-Uhlenbeck process, the mathematical expression of which is:

[0036] in, This is the noise value. This is the attenuation coefficient, with a value of 0.995. The mean is 0. The noise intensity is set to 0.1. This is the Wiener process increment. The noise has mean regression properties, which helps to achieve a balance between exploration and exploitation. In step (3), the target network is updated every 100 steps, that is, after every 100 environmental interactions, the parameters of the current policy network are copied to the target network to calculate the target Q value, reducing instability during training. In step (3), the experience replay buffer has a capacity of 10,000 sets of quadruplets, each set containing state s, action a, reward r, and next state s'. A uniform sampling strategy is used to extract small batches of samples from the buffer for training, breaking data correlation and improving training stability. The reward function in step (3) is designed as follows:

[0037] in, , , , These are the weight coefficients for each reward, corresponding to improved pulverizing efficiency, reduced unit energy consumption, decreased wear rate, and enhanced system stability, respectively. The weight values ​​are consistent with the optimization function, and are 0.4, 0.3, 0.2, and 0.1, respectively. The optimization process in step (3) is executed once every 5 seconds, and the time limit for each iteration is 200 milliseconds to ensure that control commands can be issued in a timely manner and meet the real-time requirements. The action space boundaries in step (3) are set as follows: the opening range of the primary air damper is 30% to 90%, the opening range of the hot air regulating valve is 20% to 85%, the speed range of the coal feeder is 300 rpm to 900 rpm, and the loading force range of the grinding roller is 5 kN to 15 kN. These boundary values ​​are derived from the safe operating range in the equipment nameplate parameters and long-term operation records, and are determined through historical data analysis to avoid equipment damage caused by control commands exceeding the limits.

[0038] In the above-mentioned multivariable self-optimizing control method for operating a coal mill intelligent agent, step (4) implements closed-loop feedback correction: the optimized control parameters are sent to the actuator, actual operation feedback data is collected synchronously, and the residual vector between the predicted value and the measured value is calculated. If the absolute value of the residual vector exceeds the set threshold of 0.05, the online fine-tuning mechanism of the model is triggered, and the weight parameters of the long short-term memory neural network are updated using incremental learning. The fine-tuning process uses the most recent 2000 sets of data within the sliding window as training samples. Specifically, the set threshold of the residual vector in step (4) is configured for different variables. The residual threshold for air volume is 0.04, the residual threshold for temperature is 0.06, the residual threshold for current is 0.05, the residual threshold for vibration is 0.07, the residual threshold for coal powder fineness is 0.08, and the residual threshold for inlet negative pressure is 0.05. In step (4), incremental learning employs an elastic weight solidification regularization method to retain network connection weights important for historical conditions, preventing catastrophic forgetting during new data learning. The regularization coefficient is set to 0.1. In step (4), the resource allocation deviation index is calculated by weighting spatial mismatch rate, target deviation, and execution lag coefficient, with weights of 0.4, 0.4, and 0.2, respectively. The deviation index threshold is set to 0.15. In step (4), the reinforcement learning agent employs a deep deterministic policy gradient algorithm. The state space includes the current deviation index and historical adjustment records, the action space is the fine-tuning amount of the multi-objective weight vector, and the reward function is the product of the deviation index decrease and the scheme stability. The closed-loop feedback correction process in step (4) is as follows: the optimized control parameters are processed through OPC. The UA protocol is sent to the on-site PLC controller, and the PLC executes corresponding actions, such as adjusting the damper opening and changing the coal feeder speed; at the same time, the DCS system continuously collects actual operating data, including air volume, temperature, current, vibration, coal powder fineness and inlet negative pressure, with a sampling frequency of once per second; the measured values ​​are compared with the model prediction values ​​in step (1) item by item, and the residual vector is calculated. The residual vector is a 6-dimensional vector, and each element is the absolute value of the difference between the predicted value and the measured value of the corresponding variable; if any residual value exceeds its set threshold, it is determined that the model prediction has a significant deviation, triggering the online fine-tuning mechanism; the fine-tuning process adopts the incremental learning method, extracting training samples from the most recent 2000 sets of data in the sliding window. The samples include the input sequence (length is 64) and the corresponding real output (length is 10). The same loss function and optimizer as the initial training are used, but the learning rate is reduced to 0.0001 to avoid drastic disturbances; to prevent catastrophic forgetting, the elastic weight solidification regularization method is adopted, that is, a regularization term is added to the loss function:

[0039] in, is the regularization coefficient, with a value of 0.1. For network connection weights, For the old weight, As the new weights, this regularization term penalizes weight changes that are important to historical tasks but not to current tasks, thereby preserving key knowledge. The residual thresholds in step (4) are set as follows: air volume residual threshold is 0.04, temperature residual threshold is 0.06, current residual threshold is 0.05, vibration residual threshold is 0.07, coal powder fineness residual threshold is 0.08, and inlet negative pressure residual threshold is 0.05. These thresholds are set based on the statistical distribution of historical operating data and control accuracy requirements to ensure that the model can respond to degradation in a timely manner without frequently triggering fine-tuning due to normal fluctuations.

[0040] In the above-mentioned multivariable self-optimizing control method for coal mill intelligent agent operation, step (5) realizes adaptive migration of operating conditions: establishing an operating mode library containing four typical operating conditions: low load, medium load, high load, and variable load; calculating the distance between the current operating state and the center of each mode library in real time; identifying the current operating condition category using the nearest neighbor classification criterion; and dynamically switching the corresponding optimization function weight coefficient and action space boundary to ensure that the control strategy maintains optimal performance in different load ranges. Specifically, in step (5), the operating mode library uses the K-means clustering algorithm to perform unsupervised division of historical operating condition data. The clustering features include load rate, coal grindability index, ambient temperature, and coal mill outlet temperature. The number of clusters is fixed at 4, and the mode center is represented by the mean vector of samples within each cluster. The operating condition identification process in step (5) is executed once every 10 seconds. If the identification results are consistent for three consecutive times, the operating condition switch is confirmed to avoid frequent mode jumps caused by instantaneous disturbances and ensure the continuity and robustness of the control strategy. In step (5), the visualized decision support report supports dynamic generation by quarter, year, and five-year planning cycle. The spatial distribution heat map adopts a hierarchical color method with 7 color levels. The target achievement radar chart includes 5-dimensional indicators. Risk warning prompts are classified and labeled according to three levels: high, medium, and low. The construction process of the operation mode library in step (5) is as follows: Collect historical data of coal mill operation in the past year, extract load rate (current output / rated output), coal quality Hardgrove Grindability Index (HGI), ambient temperature (°C), and coal mill outlet temperature (°C) as clustering features, with a data volume of no less than 100,000 records; adopt the K-means clustering algorithm, set the number of clusters K=4, initialize the cluster center using the K-means++ algorithm, and iterate until the square error within the cluster converges; after clustering, each cluster represents a typical working condition, named low load, medium load, high load, and variable load respectively; calculate the mean vector of the samples in each cluster as the mode center of the working condition and store it in the operation mode library. In step (5), the operating condition identification process is executed every 10 seconds. Each time, the feature vector of the current operating state is obtained, and its Euclidean distance with the four mode centers is calculated. The mode with the smallest distance is selected as the current operating condition category. If the identification results are the same for three consecutive times, the operating condition is confirmed to be switched. Otherwise, the original operating condition is maintained. This mechanism can effectively suppress misjudgments caused by instantaneous disturbances. After the operating condition is switched in step (5), the weight coefficients of the optimization function and the action space boundary are automatically adjusted. For example, under low load conditions, the weight of unit energy consumption is increased to 0.4, the weight of pulverizing efficiency is reduced to 0.3, and the action space boundary is appropriately tightened to reduce energy consumption. Under high load conditions, the weight of pulverizing efficiency is increased to 0.5, the weight of system stability is increased to 0.15, and the action space boundary is widened to ensure output and stability. Under variable load conditions, all weights remain unchanged, but smooth transition logic is added to avoid drastic fluctuations in control parameters.

[0041] In addition, it includes: establishing a coal mill health status assessment module, which calculates the equipment degradation index based on vibration spectrum characteristics and current harmonic components. When the degradation index exceeds 0.75, the weighting coefficient of the equipment wear rate in the optimization function is automatically adjusted to 0.3 to prioritize the safe operation of the equipment. Specifically, the coal mill health status assessment module consists of a signal acquisition unit, a spectrum analysis unit, and a degradation index calculation unit. The signal acquisition unit continuously acquires vibration and current signals at a sampling frequency of 10 kHz and a data length of 1 second. After anti-aliasing filtering, the data is sent to the spectrum analysis unit. The spectrum analysis unit uses Fast Fourier Transform (FFT) to calculate the spectrum of the vibration signal, extracts the dominant frequency and its harmonic components, and calculates the effective value of vibration. At the same time, it calculates the harmonic components of the current signal and extracts the amplitudes of the 3rd, 5th, and 7th harmonics. The degradation index calculation unit constructs a comprehensive evaluation model based on the effective value of vibration and the harmonic components of current. It uses fuzzy comprehensive evaluation method to normalize the effective value of vibration and the amplitude of current harmonics to the interval [0,1], sets a membership function, and calculates its membership degree to "normal", "mild degradation", "moderate degradation" and "severe degradation". Finally, the degradation index is obtained by weighted average, with a value range of [0,1], where 0 represents completely normal and 1 represents severe degradation. When the degradation index exceeds 0.75, the system automatically increases the weighting coefficient of the equipment wear rate in the optimization function from 0.2 to 0.3, prioritizing the safe operation of the equipment and preventing further deterioration.

[0042] In addition, it includes: configuring a human-machine collaborative intervention interface. When the system detects that the control parameters deviate from the operator's historical operating habits by more than 20% for five consecutive optimizations, it generates a strategy deviation alarm and provides parameter adjustment suggestions for operators to confirm or override. Specifically, the human-machine collaborative intervention interface consists of a historical operating habit database, a deviation calculation module, and an alarm generation module. The historical operating habit database stores the operator's operation records over the past year, including the time, operating conditions, and control parameter values ​​for each operation, and extracts typical operating patterns through cluster analysis. After each optimization, the deviation calculation module calculates the similarity between the current optimization result and the historical operating habits using cosine similarity calculation. If the similarity of five consecutive optimization results is less than 0.8 (i.e., the deviation exceeds 20%), an alarm is triggered. The alarm generation module generates a strategy deviation alarm, displays the difference between the current optimized parameters and historical habits, and provides parameter adjustment suggestions, such as "It is recommended to adjust the coal feeder speed from 800 rpm to 750 rpm." Operators can confirm or override the suggestion in the human-machine interface, ensuring that the system achieves a balance between automation and manual intervention.

[0043] In the above-mentioned multivariable self-optimizing control method for operating a coal mill intelligent agent, steps (1) to (5) constitute a complete closed-loop control process, and each step is tightly coupled with the data flow and control flow. Specifically, the predicted output of step (1) serves as the state input of step (3), the optimization result of step (3) serves as the control command of step (4), the feedback data of step (4) is used for model fine-tuning in step (1), and the working condition identification result of step (5) is used for parameter adjustment in steps (2) and (3), forming a multi-level, adaptive intelligent control system. The entire system is deployed on an industrial Internet of Things platform, adopts a microservice architecture, and each module runs independently and communicates asynchronously through message queues (such as Kafka) to ensure high availability and scalability of the system. The system runs on the Linux operating system, is developed using Python, and the core algorithm is implemented based on the TensorFlow and PyTorch frameworks. It is deployed on a GPU-accelerated server and supports large-scale parallel computing. The system log records all key events, including model training, parameter adjustment, working condition switching, alarm generation, etc., which facilitates later auditing and fault diagnosis.

[0044] Example 2 To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0045] Currently, coal mill operation suffers from problems such as strong multivariate coupling, frequent changes in operating conditions, and difficulty in real-time optimization of control parameters, leading to large fluctuations in pulverizing efficiency, high energy consumption, accelerated equipment wear, and insufficient system stability. To address these technical problems, this invention proposes a multivariate self-optimizing control method for operating a coal mill intelligent agent, and applies it to this method.

[0046] In the above-mentioned multivariable self-optimizing control method for coal mill intelligent operation, step (1) involves constructing a multivariable coupled dynamic model: based on historical coal mill operation data, time series data of six key variables, including air volume, temperature, current, vibration, coal powder fineness, and inlet negative pressure, are collected. The data is preprocessed using a time series segmentation method with a sliding window length of 128, and a nonlinear dynamic coupling relationship model between variables is established through a long short-term memory neural network based on an attention mechanism. The predicted values ​​of each variable in the next control cycle are then output. Specifically, in step (1), the historical operation data is sampled once per second. During the data cleaning process, time periods with a missing rate greater than 5% are removed, and outliers are corrected using the three-standard-deviation method to ensure the integrity and reliability of the input data. In step (1), the long short-term memory neural network based on an attention mechanism contains three hidden layers, each with 256 neurons, 8 attention heads, an input sequence length of 64, and an output prediction step size of 10. The model training uses a weighted loss function of mean square error and mean absolute error, with a weight ratio of 7:3. The remote sensing image data in step (1) includes high-resolution optical images, synthetic aperture radar images, and hyperspectral images, with a temporal resolution of no less than 15 days and a spatial resolution better than 2 meters. Multi-temporal image alignment is achieved through geometric correction and radiometric normalization processing based on control points. The real estate registration data in step (1) covers property rights information of natural resources such as land, forest land, water areas, and minerals. Entities are associated through unified social credit codes and spatial location codes to ensure accurate matching of property rights boundaries and geographic coordinates, with a matching error of less than 0.5 meters. In step (1), the data acquisition module consists of an industrial-grade sensor array, including an air volume sensor (model: ABB 800XA), a temperature sensor (model: PT100), a current transformer (model: LZZBJ9-10), a vibration accelerometer (model: PCB 356A15), a coal powder fineness online analyzer (model: MettlerToledo FMS-200), and an inlet negative pressure transmitter (model: Endress+Hauser DMP336). All sensors are connected to the distributed control system (DCS) and synchronized to the central database. The data transmission protocol uses Modbus TCP / IP, the communication rate is set to 10 Mbps, the data packet size is 1024 bytes, a complete state vector is generated every second, and the storage format is a CSV file. The field structure includes timestamp (accurate to milliseconds), variable name, value, unit, and quality flag (0 indicates valid, 1 indicates invalid). In step (1), the time series segmentation method with a sliding window length of 128 divides the original time series into continuous and overlapping subsequences. Each subsequence contains 128 sampling points, and the sliding step between adjacent windows is 1, forming a set of time series segments for subsequent model training and inference.In step (1), the data cleaning process first checks the integrity of the original data. If the missing rate of a certain variable exceeds 5% in 128 consecutive sampling points, the data in that time period is marked as invalid and removed from the training set. For outliers in a single sampling point, the three-standard-deviation method is used for identification and correction. That is, the mean μ and standard deviation σ of the variable in the sliding window are calculated. If the value x of a certain point satisfies |x - μ|>3σ, it is replaced with the boundary value in μ + 3σ or μ - 3σ to eliminate extreme noise interference. The long short-term memory neural network structure based on the attention mechanism in step (1) consists of an input layer, three stacked LSTM layers, an attention mechanism layer, a fully connected layer, and an output layer. The input layer receives a time sequence of length 64, with each time step containing a 6-dimensional feature vector corresponding to air volume, temperature, current, vibration, coal powder fineness, and inlet negative pressure. The input sequence is mapped to a 128-dimensional dense vector space by the embedding layer and then fed into the first LSTM layer. Each LSTM layer contains 256 memory units, using the tanh activation function and sigmoid gating mechanism to hide the state. The state and cell state are transmitted between layers. The sequence output by the third LSTM layer is used as the query and key input for the attention mechanism. There are 8 attention heads, and each head independently calculates the attention weight. After normalization by the softmax function, the weighted sum is used to obtain the context vector. This vector is then mapped to the output layer through two fully connected network layers (128 neurons per layer, ReLU activation). The output layer is a 6-dimensional vector, which corresponds to the predicted values ​​of each variable in the next control cycle. The prediction step size is 10, that is, the model can predict the state of the next 10 time steps at the same time. In step (1), the model training adopts a weighted loss function of mean squared error (MSE) and mean absolute error (MAE) with a weight ratio of 7:3. The expression of the loss function is:

[0047] in, The calculation formula is:

[0048] The formula for calculating MAE is:

[0049] For the sample size, For the true value, These are predicted values. The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. An early stopping mechanism was implemented, terminating training if the loss on the validation set did not decrease after 10 consecutive epochs to prevent overfitting. After training, the model parameters were saved in TensorFlow SavedModel format and deployed on edge computing nodes to support low-latency inference services.

[0050] In the above-mentioned multivariate self-optimization control method for operating a coal mill intelligent agent, step (2) defines a multi-objective optimization function: based on four core indicators—pulverizing efficiency, unit energy consumption, equipment wear rate, and system stability—a weighted multi-objective optimization function is constructed, where the weight coefficient for pulverizing efficiency is 0.4, the weight coefficient for unit energy consumption is 0.3, the weight coefficient for equipment wear rate is 0.2, and the weight coefficient for system stability is 0.1. Soft constraint boundaries are introduced to limit the adjustment range of air volume and temperature, ensuring that the control commands are within the safe operating range of the equipment. Specifically, in step (2), pulverizing efficiency is calculated as the ratio of qualified coal powder production to raw coal consumption per unit time; unit energy consumption is the ratio of total power consumption of the coal mill to coal powder production; equipment wear rate is evaluated based on the weighted index of the effective value of vibration and the standard deviation of current fluctuation; and system stability is measured by the reciprocal of the standard deviation of air temperature and air pressure. In step (2), the graph neural network employs a graph attention mechanism. The node feature vector dimension is 128, and the edge weights are calculated by multiplying the spatial distance decay function with the functional similarity. The decay coefficient is set to 0.8, and the similarity threshold is set to 0.7. In step (2), the spatial autocorrelation analysis uses a combination of the Moran index and the Gillley index for discrimination. The calculation window radius is set to 5 kilometers, and the significance level is set to 0.05. This is used to identify high-high and low-low clustering regions as the initial connection basis for the association graph. The multi-objective optimization function in step (2) is as follows:

[0051] in, To improve powder production efficiency, Energy consumption per unit For equipment wear rate, For the maximum permissible wear rate, For wind temperature standard deviation, For wind pressure standard deviation. Pulverizing efficiency. The calculation formula is:

[0052] in, Qualified coal powder output (unit: kg / h). Density of pulverized coal (unit: kg / m³). The qualified rate of pulverized coal (unit: %) This refers to raw coal consumption (unit: kg / h). Unit energy consumption. The calculation formula is:

[0053] in, Total power consumption of the coal mill (unit: kWh / h). Coal powder production (unit: kg / h). Equipment wear rate. The calculation formula is:

[0054] in, The effective value of vibration (unit: mm / s). Reference vibration value (unit: mm / s). The standard deviation of current fluctuation (unit: A) The reference current fluctuation standard deviation (unit: A) This is a weighting factor, with a value of 0.6. System stability is measured by the reciprocal of the standard deviations of wind temperature and wind pressure, i.e.:

[0055] in, The standard deviation of air temperature (unit: °C) The standard deviation of wind pressure is expressed in Pa. In step (2), the soft constraint boundary limits the adjustment range of air volume and temperature. The air volume adjustment range is set to 30% to 90%, and the temperature adjustment range is set to 180°C to 320°C. Control commands exceeding this range will be truncated to the boundary value to prevent the equipment from operating beyond its limits. The weight coefficients of the optimization function in step (2) can be dynamically adjusted according to the operating conditions. For example, in the low load stage, the weight of unit energy consumption can be increased to 0.4, and the weight of pulverizing efficiency can be reduced to 0.3 to prioritize reducing energy consumption; in the high load stage, the weight of pulverizing efficiency can be increased to 0.5, and the weight of system stability can be increased to 0.15 to ensure output and stability.

[0056] In the above-mentioned multivariable self-optimizing control method for operating a coal mill intelligent agent, step (3) involves performing online strategy optimization: the predicted value output in step (1) is used as the state input, combined with the current working condition label, and a deep deterministic strategy gradient algorithm is used to search for the optimal combination of control parameters in the action space. The action space has a dimension of 4, corresponding to the primary damper opening, hot air regulating valve opening, coal feeder speed, and grinding roller loading force, respectively. The optimization process is performed once every 5 seconds, and each iteration does not exceed 200 milliseconds. Specifically, the exploration noise of the deep deterministic strategy gradient algorithm in step (3) is generated using the Ornstein-Uhlenbeck process, with a noise attenuation coefficient of 0.995. The target network update cycle is 100 steps, and the experience replay buffer capacity is 10,000 sets of state-action-reward-next state quadruplets. In step (3), the action space boundary is set according to the equipment nameplate parameters and historical safe operation records. The primary air damper opening range is 30% to 90%, the hot air regulating valve opening range is 20% to 85%, the coal feeder speed range is 300 rpm to 900 rpm, and the grinding roller loading force range is 5 kN to 15 kN. In step (3), the ecological security target uses the ecological carrying capacity index as a quantitative indicator, the economic efficiency target uses the unit resource output ratio as a quantitative indicator, and the social equity target uses the inverse of the Gini coefficient as a quantitative indicator. The three constitute a three-dimensional optimization target space. In step (3), the improved multi-objective particle swarm optimization algorithm introduces an elite archiving mechanism and adaptive inertia weight. The population size is set to 200, the maximum number of iterations is set to 500, and the convergence criterion is that the Pareto front change rate is less than 1% for 50 consecutive generations. In step (3), the state space of the deep deterministic policy gradient algorithm is formed by concatenating the 6-dimensional predicted value vector output from step (1) with the current working condition label (category encoding), resulting in a 10-dimensional state vector. This vector is then input into the policy network, which is a three-layer fully connected network with 128 neurons per layer, using the ReLU activation function. The output layer is a 4-dimensional action vector corresponding to four control parameters. In step (3), the exploration noise is generated using the Ornstein-Uhlenbeck process, the mathematical expression of which is:

[0057] in, This is the noise value. This is the attenuation coefficient, with a value of 0.995. The mean is 0. The noise intensity is set to 0.1. This is the Wiener process increment. The noise has mean regression properties, which helps to achieve a balance between exploration and exploitation. In step (3), the target network is updated every 100 steps, that is, after every 100 environmental interactions, the parameters of the current policy network are copied to the target network to calculate the target Q value, reducing instability during training. In step (3), the experience replay buffer has a capacity of 10,000 sets of quadruplets, each set containing state s, action a, reward r, and next state s'. A uniform sampling strategy is used to extract small batches of samples from the buffer for training, breaking data correlation and improving training stability. The reward function in step (3) is designed as follows:

[0058] in, , , , These are the weight coefficients for each reward, corresponding to improved pulverizing efficiency, reduced unit energy consumption, decreased wear rate, and enhanced system stability, respectively. The weight values ​​are consistent with the optimization function, and are 0.4, 0.3, 0.2, and 0.1, respectively. The optimization process in step (3) is executed once every 5 seconds, and the time limit for each iteration is 200 milliseconds to ensure that control commands can be issued in a timely manner and meet the real-time requirements. The action space boundaries in step (3) are set as follows: the opening range of the primary air damper is 30% to 90%, the opening range of the hot air regulating valve is 20% to 85%, the speed range of the coal feeder is 300 rpm to 900 rpm, and the loading force range of the grinding roller is 5 kN to 15 kN. These boundary values ​​are derived from the safe operating range in the equipment nameplate parameters and long-term operation records, and are determined through historical data analysis to avoid equipment damage caused by control commands exceeding the limits.

[0059] In the above-mentioned multivariable self-optimizing control method for operating a coal mill intelligent agent, step (4) implements closed-loop feedback correction: the optimized control parameters are sent to the actuator, actual operation feedback data is collected synchronously, and the residual vector between the predicted value and the measured value is calculated. If the absolute value of the residual vector exceeds the set threshold of 0.05, the online fine-tuning mechanism of the model is triggered, and the weight parameters of the long short-term memory neural network are updated using incremental learning. The fine-tuning process uses the most recent 2000 sets of data within the sliding window as training samples. Specifically, the set threshold of the residual vector in step (4) is configured for different variables. The residual threshold for air volume is 0.04, the residual threshold for temperature is 0.06, the residual threshold for current is 0.05, the residual threshold for vibration is 0.07, the residual threshold for coal powder fineness is 0.08, and the residual threshold for inlet negative pressure is 0.05. In step (4), incremental learning employs an elastic weight solidification regularization method to retain network connection weights important for historical conditions, preventing catastrophic forgetting during new data learning. The regularization coefficient is set to 0.1. In step (4), the resource allocation deviation index is calculated by weighting spatial mismatch rate, target deviation, and execution lag coefficient, with weights of 0.4, 0.4, and 0.2, respectively. The deviation index threshold is set to 0.15. In step (4), the reinforcement learning agent employs a deep deterministic policy gradient algorithm. The state space includes the current deviation index and historical adjustment records, the action space is the fine-tuning amount of the multi-objective weight vector, and the reward function is the product of the deviation index decrease and the scheme stability. The closed-loop feedback correction process in step (4) is as follows: the optimized control parameters are processed through OPC. The UA protocol is sent to the on-site PLC controller, and the PLC executes corresponding actions, such as adjusting the damper opening and changing the coal feeder speed; at the same time, the DCS system continuously collects actual operating data, including air volume, temperature, current, vibration, coal powder fineness and inlet negative pressure, with a sampling frequency of once per second; the measured values ​​are compared with the model prediction values ​​in step (1) item by item, and the residual vector is calculated. The residual vector is a 6-dimensional vector, and each element is the absolute value of the difference between the predicted value and the measured value of the corresponding variable; if any residual value exceeds its set threshold, it is determined that the model prediction has a significant deviation, triggering the online fine-tuning mechanism; the fine-tuning process adopts the incremental learning method, extracting training samples from the most recent 2000 sets of data in the sliding window. The samples include the input sequence (length is 64) and the corresponding real output (length is 10). The same loss function and optimizer as the initial training are used, but the learning rate is reduced to 0.0001 to avoid drastic disturbances; to prevent catastrophic forgetting, the elastic weight solidification regularization method is adopted, that is, a regularization term is added to the loss function:

[0060] in, is the regularization coefficient, with a value of 0.1. For network connection weights, For the old weight, As the new weights, this regularization term penalizes weight changes that are important to historical tasks but not to current tasks, thereby preserving key knowledge. The residual thresholds in step (4) are set as follows: air volume residual threshold is 0.04, temperature residual threshold is 0.06, current residual threshold is 0.05, vibration residual threshold is 0.07, coal powder fineness residual threshold is 0.08, and inlet negative pressure residual threshold is 0.05. These thresholds are set based on the statistical distribution of historical operating data and control accuracy requirements to ensure that the model can respond to degradation in a timely manner without frequently triggering fine-tuning due to normal fluctuations.

[0061] In the above-mentioned multivariable self-optimizing control method for coal mill intelligent agent operation, step (5) realizes adaptive migration of operating conditions: establishing an operating mode library containing four typical operating conditions: low load, medium load, high load, and variable load; calculating the distance between the current operating state and the center of each mode library in real time; identifying the current operating condition category using the nearest neighbor classification criterion; and dynamically switching the corresponding optimization function weight coefficient and action space boundary to ensure that the control strategy maintains optimal performance in different load ranges. Specifically, in step (5), the operating mode library uses the K-means clustering algorithm to perform unsupervised division of historical operating condition data. The clustering features include load rate, coal grindability index, ambient temperature, and coal mill outlet temperature. The number of clusters is fixed at 4, and the mode center is represented by the mean vector of samples within each cluster. The operating condition identification process in step (5) is executed once every 10 seconds. If the identification results are consistent for three consecutive times, the operating condition switch is confirmed to avoid frequent mode jumps caused by instantaneous disturbances and ensure the continuity and robustness of the control strategy. In step (5), the visualized decision support report supports dynamic generation by quarter, year, and five-year planning cycle. The spatial distribution heat map adopts a hierarchical color method with 7 color levels. The target achievement radar chart includes 5-dimensional indicators. Risk warning prompts are classified and labeled according to three levels: high, medium, and low. The construction process of the operation mode library in step (5) is as follows: Collect historical data of coal mill operation in the past year, extract load rate (current output / rated output), coal quality Hardgrove Grindability Index (HGI), ambient temperature (°C), and coal mill outlet temperature (°C) as clustering features, with a data volume of no less than 100,000 records; adopt the K-means clustering algorithm, set the number of clusters K=4, initialize the cluster center using the K-means++ algorithm, and iterate until the square error within the cluster converges; after clustering, each cluster represents a typical working condition, named low load, medium load, high load, and variable load respectively; calculate the mean vector of the samples in each cluster as the mode center of the working condition and store it in the operation mode library. In step (5), the operating condition identification process is executed every 10 seconds. Each time, the feature vector of the current operating state is obtained, and its Euclidean distance with the four mode centers is calculated. The mode with the smallest distance is selected as the current operating condition category. If the identification results are the same for three consecutive times, the operating condition is confirmed to be switched. Otherwise, the original operating condition is maintained. This mechanism can effectively suppress misjudgments caused by instantaneous disturbances. After the operating condition is switched in step (5), the weight coefficients of the optimization function and the action space boundary are automatically adjusted. For example, under low load conditions, the weight of unit energy consumption is increased to 0.4, the weight of pulverizing efficiency is reduced to 0.3, and the action space boundary is appropriately tightened to reduce energy consumption. Under high load conditions, the weight of pulverizing efficiency is increased to 0.5, the weight of system stability is increased to 0.15, and the action space boundary is widened to ensure output and stability. Under variable load conditions, all weights remain unchanged, but smooth transition logic is added to avoid drastic fluctuations in control parameters.

[0062] In addition, it includes: establishing a coal mill health status assessment module, which calculates the equipment degradation index based on vibration spectrum characteristics and current harmonic components. When the degradation index exceeds 0.75, the weighting coefficient of the equipment wear rate in the optimization function is automatically adjusted to 0.3 to prioritize the safe operation of the equipment. Specifically, the coal mill health status assessment module consists of a signal acquisition unit, a spectrum analysis unit, and a degradation index calculation unit. The signal acquisition unit continuously acquires vibration and current signals at a sampling frequency of 10 kHz and a data length of 1 second. After anti-aliasing filtering, the data is sent to the spectrum analysis unit. The spectrum analysis unit uses Fast Fourier Transform (FFT) to calculate the spectrum of the vibration signal, extracts the dominant frequency and its harmonic components, and calculates the effective value of vibration. At the same time, it calculates the harmonic components of the current signal and extracts the amplitudes of the 3rd, 5th, and 7th harmonics. The degradation index calculation unit constructs a comprehensive evaluation model based on the effective value of vibration and the harmonic components of current. It uses fuzzy comprehensive evaluation method to normalize the effective value of vibration and the amplitude of current harmonics to the interval [0,1], sets a membership function, and calculates its membership degree to "normal", "mild degradation", "moderate degradation" and "severe degradation". Finally, the degradation index is obtained by weighted average, with a value range of [0,1], where 0 represents completely normal and 1 represents severe degradation. When the degradation index exceeds 0.75, the system automatically increases the weighting coefficient of the equipment wear rate in the optimization function from 0.2 to 0.3, prioritizing the safe operation of the equipment and preventing further deterioration.

[0063] In addition, it includes: configuring a human-machine collaborative intervention interface. When the system detects that the control parameters deviate from the operator's historical operating habits by more than 20% for five consecutive optimizations, it generates a strategy deviation alarm and provides parameter adjustment suggestions for operators to confirm or override. Specifically, the human-machine collaborative intervention interface consists of a historical operating habit database, a deviation calculation module, and an alarm generation module. The historical operating habit database stores the operator's operation records over the past year, including the time, operating conditions, and control parameter values ​​for each operation, and extracts typical operating patterns through cluster analysis. After each optimization, the deviation calculation module calculates the similarity between the current optimization result and the historical operating habits using cosine similarity calculation. If the similarity of five consecutive optimization results is less than 0.8 (i.e., the deviation exceeds 20%), an alarm is triggered. The alarm generation module generates a strategy deviation alarm, displays the difference between the current optimized parameters and historical habits, and provides parameter adjustment suggestions, such as "It is recommended to adjust the coal feeder speed from 800 rpm to 750 rpm." Operators can confirm or override the suggestion in the human-machine interface, ensuring that the system achieves a balance between automation and manual intervention.

[0064] The foregoing has shown and described 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 embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for operating a coal mill intelligent agent under multivariable self-optimization control, characterized in that: Includes the following steps: Step (1) Construct a multivariate coupled dynamic model: Based on the historical operation data of the coal mill, collect time series data of multiple key variables such as air volume, temperature, current, vibration, coal powder fineness and inlet negative pressure. Use a time series segmentation method with a sliding window length of 128 to preprocess the data, and establish a nonlinear dynamic coupling relationship model between variables through a long short-term memory neural network based on the attention mechanism, and output the predicted value of each variable in the next control cycle. Step (2) Define the multi-objective optimization function: Based on multiple core indicators such as powdering efficiency, unit energy consumption, equipment wear rate and system stability, construct a weighted multi-objective optimization function, and introduce soft constraint boundaries to limit the adjustment range of air volume and temperature to ensure that the control commands are within the safe operating range of the equipment; Step (3) Perform online strategy optimization: Take the predicted value output in step (1) as the state input, combine it with the current working condition label, and use the deep deterministic strategy gradient algorithm to search for the optimal combination of control parameters in the action space; Step (4) Implement closed-loop feedback correction: Send the optimized control parameters to the actuator, collect actual operation feedback data synchronously, calculate the residual vector between the predicted value and the measured value. If the absolute value of the residual vector exceeds the set threshold, trigger the online fine-tuning mechanism of the model, update the weight parameters of the long short-term memory neural network using incremental learning, and use the most recent 2000 sets of data in the sliding window as training samples in the fine-tuning process. Step (5) Implement adaptive migration of operating conditions: Establish an operating mode library containing four typical operating conditions: low load, medium load, high load and variable load. By calculating the distance between the current operating state and the center of each operating mode library in real time, the current operating condition category is identified by the nearest neighbor classification criterion. The corresponding optimization function weight coefficient and action space boundary are dynamically switched to ensure that the control strategy maintains optimal performance in different load ranges.

2. The method for operating a coal mill intelligent agent with multivariable self-optimization control according to claim 1, characterized in that: In step (1), the historical running data is sampled once per second. The data cleaning process removes time periods with a missing rate greater than 5%, and outliers are corrected using the three-standard-deviation method to ensure the integrity and reliability of the input data. The long short-term memory neural network based on the attention mechanism contains 3 hidden layers, each with 256 neurons, 8 attention heads, an input sequence length of 64, and an output prediction step size of 10. The model training uses a weighted loss function of mean squared error and mean absolute error, with a weight ratio of 7 to 3.

3. The method for operating a coal mill intelligent agent under multivariable self-optimization control according to claim 1, characterized in that: In step (2), the pulverizing efficiency is calculated by the ratio of qualified coal powder output to raw coal consumption per unit time. The unit energy consumption is the ratio of total power consumption of the coal mill to coal powder output. The equipment wear rate is evaluated by the weighted index of the effective value of vibration and the standard deviation of current fluctuation. The system stability is measured by the reciprocal of the standard deviation of air temperature and air pressure. The soft constraint boundary limits the adjustment range of air volume and temperature. The air volume adjustment range is set to 30% to 90%, and the temperature adjustment range is set to 180°C to 320°C.

4. The method for operating a coal mill intelligent agent under multivariable self-optimization control according to claim 1, characterized in that: In step (3), the exploration noise of the deep deterministic strategy gradient algorithm is generated using the Ornstein-Uhlenbeck process, with a noise attenuation coefficient of 0.

995. The target network update cycle is 100 steps, and the experience replay buffer capacity is 10,000 sets of state-action-reward-next state quadruplets. The action space boundary is set according to the equipment nameplate parameters and historical safe operation records. The primary air damper opening range is 30% to 90%, the hot air regulating valve opening range is 20% to 85%, the coal feeder speed range is 300 rpm to 900 rpm, and the grinding roller loading force range is 5 kN to 15 kN.

5. The method for operating a coal mill intelligent agent under multivariable self-optimization control according to claim 1, characterized in that: In step (4), the threshold values ​​for the residual vectors are configured for different variables: the residual threshold for air volume is 0.04, the residual threshold for temperature is 0.06, the residual threshold for current is 0.05, the residual threshold for vibration is 0.07, the residual threshold for coal powder fineness is 0.08, and the residual threshold for inlet negative pressure is 0.

05. The incremental learning adopts the elastic weight solidification regularization method to retain the network connection weights that are important for historical working conditions and prevent the model from catastrophic forgetting during the learning process of new data. The regularization coefficient is set to 0.

1.

6. The method for operating a coal mill intelligent agent under multivariable self-optimization control according to claim 1, characterized in that: In step (5), the operating mode library performs unsupervised division of historical operating condition data using the K-means clustering algorithm. The clustering features include load rate, coal grindability index, ambient temperature and coal mill outlet temperature. The number of clusters is fixed at 4, and the mode center is represented by the mean vector of the samples in each cluster. The operating condition identification process is executed once every 10 seconds. If the identification results are consistent for 3 consecutive times, the operating condition switch is confirmed to avoid frequent mode jumps caused by instantaneous disturbances.

7. The method for operating a coal mill intelligent agent under multivariable self-optimization control according to claim 1, characterized in that: It also includes establishing a coal mill health status assessment module, which calculates the equipment deterioration index based on vibration spectrum characteristics and current harmonic components. When the deterioration index exceeds 0.75, the weight coefficient of the equipment wear rate in the optimization function is automatically adjusted to 0.3 to prioritize the safe operation of the equipment. The equipment deterioration index is calculated by fuzzy comprehensive evaluation method, which is obtained by normalizing the effective value of vibration and the amplitude of current harmonics and then taking a weighted average. The value range is 0 to 1.

8. The method for operating a coal mill intelligent agent under multivariable self-optimization control according to claim 7, characterized in that: The vibration spectrum features are extracted by fast Fourier transform to obtain the main frequency and its harmonic components, and the current harmonic components are extracted by extracting the amplitudes of the 3rd, 5th and 7th harmonics. The signal acquisition unit has a sampling frequency of 10 kHz and a data length of 1 second. After anti-aliasing filtering, the data is sent to the spectrum analysis unit for processing.

9. The method for operating a coal mill intelligent agent under multivariable self-optimization control according to claim 1, characterized in that: It also includes a human-machine collaborative intervention interface. When the system detects that the optimization results of the control parameters deviate from the operator's historical operating habits by more than 20% for five consecutive times, it generates a strategy deviation alarm and provides parameter adjustment suggestions for operators to confirm or override. The deviation is calculated using cosine similarity, and the historical operating habit database extracts typical operating patterns through cluster analysis.

10. A method for operating a coal mill intelligent agent under multivariable self-optimization control according to claim 9, characterized in that: When generating an alarm, the human-machine collaborative intervention interface displays the difference between the current optimized parameters and historical habits, and provides specific parameter adjustment suggestions. Operators can confirm or override these suggestions through the human-machine interface. The historical operation habit database contains records of time, operating conditions, and control parameter values, with a storage period of no less than one year, and is used to build an operation habit benchmark model.

Images