Furnace temperature control system

Abstract

The application discloses a furnace temperature control system, and belongs to the technical field of industrial automation control, which comprises the following steps: real-time monitoring and collecting multi-dimensional data of a physical furnace body, preprocessing, generating a state vector, setting a working condition discrimination method, classifying various working conditions, outputting working condition labels and characteristic parameters, establishing a health degree evaluation model, real-time monitoring of an actuator state, preventing frequent maintenance and shortening of service life caused by overuse of equipment, constructing a furnace virtual model, setting a real-time prediction method, simulating furnace temperature distribution to predict future temperature trends, determining whether to start a rehearsal mode, outputting an optimal scheme, realizing early intervention of prediction and decision-making, reading working condition recognition and temperature prediction results, setting a decision engine method, generating dynamic control instructions and executing, comparing data of the physical furnace body and the digital twin, setting a self-evolution method, online calibration of the virtual model, and correction.

Description

Technical Field

[0001] This invention relates to a furnace body temperature control system, belonging to the field of industrial automation control technology. Background Technology

[0002] In the industrial production field, the furnace body is a key piece of equipment, and the accuracy and stability of its temperature control directly affect product quality, production efficiency and energy consumption. With the continuous improvement of industrial automation, the requirements for furnace body temperature control are also getting higher and higher.

[0003] With the popularization of flexible manufacturing models, actual production often faces complex dynamic scenarios such as frequent start-ups and shutdowns, sudden load changes, and multi-furnace coordination. Examples include batch cleaning and material changes required for chemical intermittent reactors, frequent start-ups and shutdowns of metallurgical electric furnaces due to process switching or peak-valley electricity pricing strategies, and sudden load changes in building material rotary kilns due to batch feeding. These scenarios challenge the dynamic response capability of temperature control. Traditional systems often only conduct simple condition tests such as steady-state heating during the design phase, lacking simulation verification of real-world complex scenarios, leading to defects exposed during actual operation. When furnaces frequently start and stop or experience sudden load changes, the nonlinearity, time-varying nature, and large hysteresis of temperature characteristics make it difficult for traditional control algorithms to adjust in a timely manner, easily resulting in temperature overshoot and control oscillations. Frequent adjustments to heating or cooling equipment not only lead to soaring energy consumption but also affect product consistency due to temperature fluctuations. More importantly, traditional control strategies, in an effort to avoid large temperature fluctuations, often excessively pursue stability by employing fixed-parameter, strong-feedback regulation. This, in turn, causes frequent actuator movements, exacerbating mechanical wear and tear, shortening lifespan, and increasing maintenance costs. This dilemma essentially reflects the fundamental limitations of traditional control technology in adapting to complex operating conditions. It cannot accurately respond to real-time changes in dynamic scenarios, and its rigid control logic leads to an imbalance between stability and equipment reliability. There is an urgent need for more intelligent control strategies and testing and verification systems to overcome these bottlenecks and achieve synergistic optimization of temperature control accuracy, energy efficiency, and equipment lifespan. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a furnace temperature control system that integrates multi-dimensional data from multiple types of sensors, preprocesses it using edge computing, uses digital twins to predict temperature trends and predictive strategies, combines operating condition-adaptive control strategies, and optimizes module iterative calibration models and strategies to solve problems such as insufficient data fusion, control lag, and dynamic coordination.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The furnace body temperature control system includes:

[0007] Real-time monitoring and acquisition of multi-dimensional data of the physical furnace body, preprocessing to generate state vectors, setting working condition discrimination methods, classifying various working conditions, establishing a health assessment model, and real-time monitoring of actuator status;

[0008] Construct a virtual model of the furnace body, set up a real-time prediction method, simulate the temperature distribution of the furnace body to predict future temperature trends, determine whether to start the pre-simulation mode, and output the optimal solution.

[0009] Read the operating condition identification and temperature prediction results, set the decision engine method, generate dynamic control commands, and execute them;

[0010] By comparing the data of the physical furnace body and the digital twin, a self-evolution method is set up to calibrate and correct the virtual model online.

[0011] Specifically, the steps for generating the state vector include:

[0012] Sensors were deployed at key locations on the furnace body to form a multi-dimensional monitoring network, and the sensor communication links were tested.

[0013] Various sensors synchronously collect data at preset frequencies, including temperature arrays, real-time load mass, and time-domain vibration sequences;

[0014] A unified reference clock is used to calibrate the sensor's time, delay compensation is performed on network transmission data, and synchronized data is output.

[0015] Adaptive Kalman filtering is used to reduce noise in temperature data, and frequency domain transformation is performed on vibration signals to extract key parameters;

[0016] The pre-processed temperature, load mass, equipment vibration status, and actuator operating status data are integrated into a unified format state vector, and after validity verification, they are transmitted through an industrial communication protocol.

[0017] Specifically, the working condition discrimination method includes:

[0018] Read the temperature data from the state vector, set a fixed time window, and calculate the rate of temperature change within the fixed time window. and temperature gradient ;

[0019] Calculate the absolute magnitude of load change by combining real-time load quality with rated load value. and relative rate of change ;

[0020] Define start / stop events, count the number of device start / stop cycles and power adjustment range, and count the number of start / stop cycles within the fixed time window in real time. Calculate start-stop interval ;

[0021] Set a threshold for the number of start / stop cycles. and start / stop interval threshold Determine if frequent start-stop conditions occur; if and If a scenario is marked as a frequent start-stop scenario, the operating condition label is 1; otherwise, it is not marked.

[0022] Set the load change threshold to Temperature response threshold is Determine if the load changes suddenly; and If the load change is sudden, the condition is marked as 2; otherwise, it is not marked.

[0023] Calculate the correlation coefficient of adjacent furnace body temperatures and temperature difference And set relevant thresholds. and temperature difference threshold To identify collaborative working conditions; if and If the scenario is a multi-furnace collaborative scenario, the operating condition label is 3; otherwise, no labeling is performed.

[0024] When the conditions for frequent start-stop scenarios, sudden load changes, and multi-furnace collaborative scenarios are not met, the scenario is judged as a steady-state operation scenario, and the operating condition label is 0.

[0025] Specifically, the working condition discrimination method further includes:

[0026] A three-layer health assessment model is constructed based on vibration signal, motor current harmonic distortion rate, and the ratio of cumulative start-stop times to rated start-stop times.

[0027] The health index of the equipment is generated using a weighted summation method, and the health level of the actuator and the corresponding response measures are output according to the set health grading rules.

[0028] The operating condition classification results, dynamic feature parameters, and equipment health status are integrated to generate an identification vector, which is then logically verified.

[0029] Specifically, in the working condition discrimination method, multiple thresholds are set, and an adaptive adjustment mechanism is set for the thresholds;

[0030] The adaptive adjustment mechanism includes:

[0031] Collect historical data on equipment operation, extract feature parameter values, and calculate statistics as the basis for threshold initialization;

[0032] Set initial thresholds for each threshold and save them to the constructed threshold benchmark library;

[0033] Data is collected in real time using a sliding window, the oldest data is removed and new data is added, and the statistics of each feature parameter within the sliding window are calculated.

[0034] Calculate the deviation rate between the feature parameters and the initial threshold. And set the deviation threshold. To determine whether to perform adaptive adjustments;

[0035] when If the initial threshold is adjusted, the adjusted threshold is filtered using an exponential smoothing algorithm; otherwise, the original initial threshold is maintained.

[0036] Identify overlapping scenarios; if there is a conflict in the threshold adjustment direction in the overlapping scenario, use the Pareto optimization algorithm to generate a compromise threshold and update the threshold of the working condition discrimination method in real time.

[0037] Verify the reasonableness of the threshold, and choose to use or roll back the threshold based on the verification results.

[0038] Specifically, the steps for constructing a virtual model of the furnace body include:

[0039] The furnace structure data is acquired using a 3D laser scanner, and point clouds are generated and noise is reduced.

[0040] The furnace body was divided into multiple finite element mesh elements using finite element software, functional areas were defined, key nodes were marked, and the thermal properties parameters and boundary conditions of the furnace body material were input to construct a three-dimensional geometric model.

[0041] Collect historical operating condition data and select key data points under typical operating conditions as calibration datasets;

[0042] The thermophysical parameters were calibrated using a Bayesian optimization algorithm, and a virtual model was output after data verification.

[0043] Specifically, the real-time prediction method includes:

[0044] The real-time state vector data is converted into input parameters for the virtual model, and the timestamps are synchronized.

[0045] The temperature field is calculated by solving the heat conduction equation using the finite element method, and corrections are made for the presence of phase transition processes to generate the temperature field matrix.

[0046] Based on the temperature field matrix, a prediction curve is generated by extrapolation, and the extreme points of the prediction curve and the time to reach the target temperature are calculated.

[0047] In non-steady-state operation scenarios, historical data is combined to generate long-term prediction curves using temporal convolutional networks, and prediction results at multiple time scales are output.

[0048] Candidate control strategies are generated based on operating conditions and equipment status, and pre-simulated using the virtual model, with pre-simulation indicators recorded.

[0049] Based on a multi-objective optimization algorithm, the reward value of the candidate control strategy is calculated, and the strategy with the largest reward value is selected as the pre-action plan.

[0050] The prediction error is calculated in real time. When the prediction error exceeds a threshold, the virtual model parameters are optimized and the virtual model is updated using a differential evolution algorithm.

[0051] Specifically, the decision engine method includes:

[0052] Construct a working condition-strategy mapping table and establish the correspondence between working condition labels and control algorithms;

[0053] The corresponding control algorithm is invoked based on the operating condition label and equipment health status, and initial parameters are loaded.

[0054] Based on the characteristics of the working conditions, the weight coefficients of temperature accuracy, energy consumption, equipment operation frequency, and equipment health indicators are dynamically adjusted to construct a multi-objective reward function and define constraints.

[0055] Using the corresponding control algorithm, calculate the control quantity for different operating conditions;

[0056] Depending on the state of the device, the control quantity is softened or made safer.

[0057] The control quantity is converted into a standardized signal and an execution attribute is attached;

[0058] Verify whether the control quantity is within the device safety threshold. If it exceeds the device safety threshold, automatically adjust it to the safety threshold and record a warning.

[0059] In a multi-furnace collaborative scenario, once the total power exceeds the grid capacity, power is dynamically allocated according to priority to handle power conflicts.

[0060] Record control process data, calculate the deviation between control commands and actual execution, score the control effect, generate feedback correction signals based on the scores, and if correction is triggered three times consecutively, revert to the effective control of the previous version.

[0061] Specifically, the self-evolution method includes:

[0062] Acquire temperature prediction error data, analyze the error distribution of key nodes, identify areas where temperature prediction deviation exceeds the standard, and generate an error heat map;

[0063] The thermal property parameters of the virtual model are optimized using the differential evolution algorithm;

[0064] Extract complete control cycle data and preprocess it to form a training set;

[0065] For a single furnace, the PPO algorithm is used to optimize control actions with temperature field, load and equipment health as the state space. For multiple furnaces, multi-agent reinforcement learning combined with graph neural network is used to model the cooperative relationship between furnaces, minimize global temperature deviation and total energy consumption, and optimize control actions.

[0066] The newly trained control strategy is validated offline in the virtual model, historical fault conditions are simulated, the control effects of the new and old control strategies are compared, and the results are updated in real time.

[0067] Temperature accuracy, energy consumption, and equipment operating frequency are defined as optimization objectives. The NSGA-II algorithm is used to generate Pareto optimal solution sets, and the weights are dynamically adjusted.

[0068] When an undefined abnormal operating condition is detected, relevant data is automatically collected, abnormal features are identified through the isolated forest algorithm, new operating condition labels are generated, and control strategies are updated.

[0069] Regularly analyze key performance indicators, use time series analysis models to predict performance trends, and optimize accordingly.

[0070] The beneficial effects of this invention are:

[0071] By deploying multiple types of sensors to build a multi-dimensional monitoring network and combining edge computing to achieve spatiotemporal registration and noise reduction of multi-source data, the problems of easy interference and transmission lag in traditional single-sensor data are solved, providing reliable data for precise control. Based on a multi-layer feature recognition system, operating conditions are accurately classified and equipment health is assessed to avoid excessive equipment wear. At the same time, digital twins are used to simulate the temperature field in real time, predict trends, and pre-play control strategies to solve the problem of large lag control delay and improve control accuracy. Dynamic control commands are generated by combining operating conditions and prediction results, taking into account multi-objective optimization, extending equipment life and reducing energy consumption. Through online calibration and strategy iteration, it adapts to time-varying factors such as furnace aging, achieving efficient and stable temperature control as a whole, improving equipment reliability and resource utilization, and contributing to the goal of circular economy. Attached Figure Description

[0072] Figure 1 This is a structural diagram of the furnace body temperature control system;

[0073] Figure 2 This is a flowchart of the working condition discrimination method of the present invention;

[0074] Figure 3 This is a flowchart of the adaptive adjustment mechanism of the present invention;

[0075] Figure 4 This is a flowchart of the real-time prediction method of the present invention. Detailed Implementation

[0076] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0077] Reference Figures 1 to 4 This embodiment describes a furnace temperature control system, including: a temperature monitoring module, an identification module, a prediction module, a decision-making module, and an optimization module;

[0078] The temperature monitoring module is used to monitor in real time through high-temperature resistant sensors deployed in key parts of the furnace body, and to collect multi-dimensional data of the physical furnace body, such as temperature, load, and equipment status. Edge computing is used for fusion processing to solve the problems of traditional single sensor data being susceptible to interference and transmission lag. At the same time, a unified timestamp synchronization mechanism is built to perform spatiotemporal registration and preprocessing of multi-dimensional data, and to reduce data noise using filtering algorithms to generate standardized state vectors, providing accurate and timely data support for subsequent modules and avoiding temperature control inaccuracies caused by data errors.

[0079] The identification module is used to acquire real-time data to identify the current operating conditions, set the operating condition discrimination method, and build a multi-layer feature recognition system to achieve accurate classification of various operating conditions, such as frequent start-stop, sudden load changes, multi-furnace coordination, and steady-state operation scenarios. It outputs operating condition labels and feature parameters to drive control strategy adaptation. At the same time, it establishes a health assessment model to monitor the status of actuators in real time. Once the equipment approaches the wear threshold, it sends adjustment instructions to the decision module to adjust the control strategy in a timely manner, preventing frequent maintenance and shortened lifespan caused by overuse of equipment. This solves the problem of excessive wear caused by ignoring equipment status.

[0080] The prediction module is used to construct a virtual model of the furnace body using digital twin technology, set up a real-time prediction method, use real-time synchronized data of the physical furnace body and simulate the temperature distribution of the furnace body to predict future temperature trends, solve the problem of temperature overshoot and oscillation caused by large lag control delay, and make advance judgments on changes in operating conditions based on the operating condition labels output by the identification module, determine whether to start the pre-run mode, and pre-run multiple control strategies in the virtual space to select the optimal solution, so as to realize early intervention in prediction and decision-making, so as to respond to temperature changes in advance and improve control accuracy.

[0081] The decision module is used to read the results of operating condition identification and temperature prediction, set the decision engine method, and use reinforcement learning algorithms to generate dynamic control commands. It takes into account temperature accuracy, energy efficiency and equipment life, and solves the problem that traditional fixed parameter control is difficult to balance multiple objectives. At the same time, by flexibly executing control commands, it avoids the actuator from frequent actions caused by strong feedback adjustment. It adopts smooth power regulation and predictive cooling control to extend equipment life, reduce energy consumption, and achieve efficient and stable temperature control.

[0082] The optimization module is used to compare the data of the physical furnace body and the digital twin, set up a self-evolution method, perform online calibration of the virtual model, and iterate the reinforcement learning strategy to correct the model deviation caused by furnace aging and material changes, thus solving the problem that the traditional model is fixed and the strategy cannot adapt to time-varying factors.

[0083] In this embodiment, multidimensional data is first collected in real time. After edge computing preprocessing, a standardized state vector is generated. Based on a multi-layer feature recognition system, the real-time data is analyzed to identify the current operating condition and assess the health of the equipment. At the same time, the physical furnace body data is synchronized in real time using a digital twin. The temperature field distribution is simulated through finite element analysis to predict future temperature trends and generate pre-action plans. Control commands are generated by combining the operating condition information and prediction results to drive the actuators. Finally, during operation, the model parameters are calibrated daily by comparing the physical furnace body data and the digital twin data. Historical data is used periodically to iteratively train the reinforcement learning strategy to continuously optimize the control performance.

[0084] Specifically, the steps for generating a standardized state vector include:

[0085] Multiple types of sensors are deployed at key locations on the furnace body, including thermocouples for temperature acquisition, microwave radar for non-contact load detection, and vibration sensors for monitoring equipment status. The sensor installation locations are determined according to the furnace structure and process requirements to ensure coverage of key areas of temperature field, load changes, and equipment operation, forming a multi-dimensional monitoring network. After the hardware installation is completed, the sensor communication links are tested to ensure stable data transmission. Key locations include the heating zone, material contact surface, cooling channel, and actuator surface.

[0086] Various sensors synchronously collect data at preset frequencies. Thermocouples continuously acquire temperature signals from different locations within the furnace body, converting them into temperature values ​​via an analog-to-digital converter. A polling mechanism is used to synchronously collect all thermocouple data, generating a temperature array. Microwave radar measures the material height, and combined with furnace geometry and material characteristics, indirectly calculates the real-time load mass. Vibration sensors collect actuator vibration signals, obtaining a time-domain vibration sequence for subsequent equipment health analysis. The expressions are shown below:

[0087] ;

[0088] In the formula, For real-time load quality, For material height, The cross-sectional area of ​​the furnace body. Bulk density of the material;

[0089] Using a unified reference clock, such as GPS or NTP server, all sensors are time-calibrated to ensure that the data acquisition timestamp is consistent with the reference time. At the same time, delay compensation is performed on network transmission data to correct time deviations caused by signal transmission, ensuring that temperature, load, and equipment status data are accurately aligned in the time dimension. This provides a time-accurate data source for subsequent operating condition analysis and outputs synchronized data with precise timestamps.

[0090] The collected raw data is denoised and features are extracted. For temperature data, an adaptive Kalman filter is used to denoise the temperature sequence, filter out high-frequency interference and abnormal fluctuations, smooth the temperature curve, and retain the true temperature change trend. For vibration signals, the time-domain vibration signal is converted into frequency-domain features, key parameters are extracted, and abnormal characteristics of equipment vibration are highlighted to facilitate subsequent equipment fault identification.

[0091] The preprocessed multi-source data is integrated into a unified state vector, which includes temperature characteristics, load quality, equipment vibration status and actuator working status. The data is validated, outliers are repaired or removed, and high-quality structured data is generated. This data is then transmitted to subsequent modules via industrial communication protocols, providing standardized input for condition identification and control decisions.

[0092] Specifically, the operating condition identification methods include:

[0093] Read temperature data from the state vector Set a fixed time window and calculate the rate of temperature change within the fixed time window. and temperature gradient The size of the fixed time window is The rate of temperature change is calculated by the ratio of the temperature difference between the first and last moments within the window to the window size, reflecting the trend of dynamic temperature change. The temperature gradient is obtained by calculating the difference between the highest and lowest temperatures within the window, characterizing the uniformity of the furnace temperature field. Dynamic features capture the real-time evolution of the temperature field, avoiding misjudgments caused by static thresholds. The expression is shown below:

[0094]

[0095]

[0096] In the formula, for The temperature value at any given moment also represents the temperature value at the last moment within the time window. , Temperature data The maximum and minimum values ​​in;

[0097] Combined with real-time load quality With rated load value Calculate the absolute magnitude of load change per unit time. and relative rate of change The absolute amplitude is the absolute value of the load quality difference between the first and last moments within the window, reflecting the absolute change in load per unit time. The relative change rate is the ratio of the absolute amplitude to the rated load value, used for cross-operating condition comparisons. By combining absolute and relative values, the degree of load fluctuation is accurately quantified; the expression is as follows:

[0098]

[0099]

[0100] In the formula, for The load quality at a given moment also represents the load quality at the last moment within the time window.

[0101] Start-stop events are defined as follows: when the power increases from 0% to more than 20% of the rated power, it is considered a start-up state; when the power decreases from more than 20% of the rated power to 0%, it is considered a stop state. Based on these start-stop events, the power adjustment signals and start-stop status of the heating and cooling equipment are collected. The number of start-stop cycles and the power adjustment range per unit time are statistically analyzed to reflect the operating status and adjustment frequency of the equipment. This is used to identify frequent start-stop conditions and assess the degree of equipment wear. The power adjustment range is the power difference between the first and last moments within the time window.

[0102] Based on equipment start-up and stop time records, the number of start-ups and stoppages within a fixed time window is counted in real time. Calculate the start and stop intervals within the window. And set a threshold for the number of start / stop cycles. and start / stop interval threshold To determine frequent start-stop conditions; start-stop interval The calculation formula is:

[0103]

[0104] In the formula, The size of the fixed time window;

[0105] like and If the condition is positive, it indicates that the device starts and stops frequently within this time window, and is marked as a frequent start-stop scenario with a working condition label of 1; otherwise, it is not marked.

[0106] Set the load change threshold to Temperature response threshold is By combining the relative rate of change of load and the rate of change of temperature, a joint judgment is made to identify sudden load changes caused by sudden changes in material input or adjustments to process parameters.

[0107] like and If the load changes suddenly within this time window, it is marked as a load change scenario and the operating condition label is 2; otherwise, it is not marked.

[0108] Calculate the correlation coefficient of adjacent furnace body temperatures and temperature difference And set relevant thresholds. and temperature difference threshold To identify collaborative working conditions; the expression is as follows:

[0109]

[0110]

[0111] In the formula, , These are the numbers of two adjacent furnace bodies. Furnace body , Covariance between , To separate the furnace bodies , standard deviation , To separate the furnace bodies , The real-time temperature value;

[0112] like and If the condition is true, it indicates that multiple furnaces are working together within this time window, and it is marked as a multi-furnace collaborative scenario with a working condition label of 3; otherwise, it is not marked.

[0113] When the real-time characteristics do not meet the judgment conditions for frequent start-stop scenarios, sudden load change scenarios, and multi-furnace collaborative scenarios, that is, no marking is made, it is automatically judged as a steady-state operation scenario, and the operating condition label is defined as 0.

[0114] The effective value of vibration is calculated based on vibration signals and defined as a basic index. It reflects the wear and tear of mechanical parts, calculates the harmonic distortion rate of motor current, and is defined as an electrical index. This reflects abnormal motor load. The ratio of the cumulative number of start-stop cycles to the rated number of start-stop cycles is defined as the lifespan index. The health status of the equipment is comprehensively evaluated through three dimensions to construct a three-layer health assessment model, avoiding the one-sidedness of assessment by a single indicator;

[0115] Use weighted summation method to generate equipment health index By rationally allocating weights, highlighting the dominant role of mechanical condition while also considering electrical performance and lifespan loss, a quantitative assessment of equipment health status is achieved; the expression is shown below:

[0116]

[0117] In the formula, , , These are the weighting coefficients for each indicator, and This can be set by those skilled in the art; in this embodiment, it is set... , , ;

[0118] Based on the set health grading rules, output the current health level of the actuator and the corresponding response measures; wherein, the health grading rules include: setting a secondary health threshold. , ,and ,when The actuator is running normally, and is in a healthy state at this time. This indicates that the actuator requires maintenance, such as lubrication or tightening. This is a warning state. The system generates control adjustments, triggers the adjustment of control strategies, and pushes maintenance work orders; at this point, the system is in a fault state.

[0119] The operating condition classification results, dynamic feature parameters, and equipment health status are integrated into a structured identification vector. The operating condition classification results are labeled as operating condition 0, 1, 2, and 3. The dynamic feature parameters include the number of start-stop cycles, start-stop interval, relative change rate of load, temperature change rate, and correlation coefficient in multi-furnace collaborative scenarios (set to 0 in non-collaborative scenarios).

[0120] Logical verification is performed on the integrated recognition vector to ensure that the working condition label is consistent with the feature parameters and that the equipment health status matches the operation frequency. Through the data verification mechanism, the reliability of the recognition results is improved, and erroneous decisions caused by sensor anomalies or rule conflicts are avoided.

[0121] Specifically, the operating condition discrimination method sets multiple thresholds, including a start / stop count threshold. Start-stop interval threshold Load change threshold Temperature response threshold Relevant thresholds and temperature difference threshold An adaptive adjustment mechanism is set for the threshold.

[0122] The adaptive adjustment mechanism includes:

[0123] Collect historical data on equipment operation, including at least three complete production cycles, covering various scenarios such as stable operation, frequent start-stop, sudden load changes, and multi-furnace coordination. Extract the corresponding characteristic parameter values ​​under each operating condition and calculate the corresponding statistics, including mean, standard deviation, and quantiles, as the basis for threshold initialization.

[0124] Add the average of historical start and stop times The initial start-stop frequency threshold was set at 1 / 3 standard deviation; the initial start-stop interval threshold was set at the average interval between frequent start-stop cases in historical data; the initial load change threshold was set at the 90th quantile of the rate of change in historical load abrupt change cases; the initial temperature response threshold was set at the median peak value of the corresponding temperature change rate; and the initial correlation threshold was set at the mean Pearson correlation coefficient of adjacent furnace temperature series in historical data. The temperature standard deviation is used as the initial temperature difference threshold, and the calculated initial threshold is saved to the constructed threshold benchmark library;

[0125] A 12-hour sliding time window is used to collect the latest operating data in real time. Every 30 minutes, the earliest 30 minutes of data are removed and new data is added to keep the data in the window the latest 12 hours of operation records. The data in the window is preprocessed and the statistics of each characteristic parameter are calculated for threshold deviation analysis.

[0126] The deviation rate between the feature parameters and the initial threshold is obtained by dividing the difference between the mean of each parameter within the latest sliding time window and the initial threshold by the initial threshold. And set the deviation threshold. To determine whether adaptive adjustment should be performed; wherein, the deviation threshold is set by those skilled in the art, and in this embodiment, it is taken as... ;

[0127] when When the threshold adaptive adjustment process is triggered, the start / stop frequency threshold, start / stop interval threshold, and related thresholds are adjusted proportionally based on the deviation rate to avoid oversensitivity or lag. The load change threshold is dynamically corrected based on the standard deviation of the current load change rate to adapt to fluctuations in material input. The temperature response threshold is adjusted according to the peak trend of the temperature change rate, and the temperature difference threshold is dynamically adjusted according to the process mode. Otherwise, the original thresholds are maintained. The expressions are as follows:

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134] In the formula, , , , , , The adjustment coefficient is determined by those skilled in the art. For continuous occurrences in scenarios of sudden load changes The number of windows, The standard deviation of the current window load change rate. This represents the average temperature difference of the furnace body at the current window. , To adjust the threshold for the number of start-stop cycles before and after, , To adjust the start-stop interval threshold before and after, , To adjust the load change threshold before and after, , To adjust the temperature response thresholds before and after, , To adjust the relevant thresholds before and after, , To adjust the temperature difference threshold before and after;

[0135] The adjusted threshold is filtered using an exponential smoothing algorithm to reduce the impact of occasional outliers and make the threshold change smooth.

[0136] Identify cross-scenarios where two or more operating conditions are triggered simultaneously, such as the coexistence of frequent start-stop and sudden load changes. Extract the distribution of feature parameters in such scenarios and analyze whether there are conflicts in the adjustment direction of each threshold in the cross-scenarios. For example, if the threshold for the number of start-stops increases while the threshold for the load change rate decreases, use the Pareto optimization algorithm to generate a compromise threshold, balance the identification priority of different operating conditions, and minimize the sum of the false positive rate and the false negative rate.

[0137] The adjusted threshold is synchronously saved to the threshold benchmark library and replaces the original threshold parameters in the working condition identification process. At the same time, the adjusted threshold parameters are imported into the digital twin model to simulate historical fault conditions, such as start-stop cases that have caused temperature overshoot, and the working condition identification accuracy is calculated to verify the rationality of the threshold.

[0138] If the simulation results show that the recognition accuracy is lower than the preset standard, the threshold adjustment rollback mechanism is triggered to restore the previous effective threshold; otherwise, the adjusted threshold is used.

[0139] The system provides real-time statistics on the control effect after identifying operating conditions during actual operation, such as temperature overshoot and equipment operation frequency. If five or more misjudgments occur consecutively, the system automatically reverts to the historical effective threshold; otherwise, the adjusted threshold is used.

[0140] Specifically, the steps for constructing a virtual model of the furnace body include:

[0141] A high-precision 3D laser scanner is used to scan the furnace body to obtain detailed structural data of the furnace body, including the external contour, heating electrode position, and cooling channel layout. Point cloud data is generated, and noise reduction algorithms, such as moving least squares, are used to remove outliers.

[0142] The furnace body is divided into multiple finite element mesh elements using finite element software. Functional areas such as heating zone, material chamber, and cooling layer are defined, key nodes are marked, and the thermal properties of the furnace body material and boundary conditions such as ambient temperature and convective heat transfer coefficient are input to construct the initial three-dimensional geometric model.

[0143] Collect historical operating condition data and select key data points under typical operating conditions as calibration datasets;

[0144] The Bayesian optimization algorithm is used to adjust the thermal property parameters of the three-dimensional geometric model to minimize the error between the predicted temperature and the actual temperature.

[0145] The model's prediction accuracy is verified using data that has not been calibrated, ensuring that the model accurately reflects the actual thermal characteristics of the furnace body, in order to output the final virtual model of the furnace body.

[0146] Specifically, real-time prediction methods include:

[0147] The data in the real-time state vector is transformed into input parameters for the virtual furnace model, such as converting heating power into heat source density and correcting the equivalent specific heat capacity of the material cavity according to the load mass. At the same time, the timestamp of the data is adjusted to align with the calculation time step of the virtual model, so as to achieve state synchronization between the physical furnace and the digital twin.

[0148] Based on real-time input parameters, the heat conduction equation is solved using the finite element method to calculate the furnace temperature field in real time and generate a nodal temperature matrix. At the same time, the implicit time integration method is used to process the time step to ensure calculation accuracy and efficiency.

[0149] If a phase change process exists in the material cavity, a latent heat correction term is introduced to simulate the effect of the phase change on the temperature field. GPU parallel computing is used to accelerate node matrix operations, realize the real-time simulation of the temperature field distribution, and generate the final temperature field matrix.

[0150] Based on the real-time temperature field matrix, the temperature of key nodes is extracted, and the second-by-second prediction curve for the next 10 minutes is generated by extrapolation to make short-term predictions and calculate the extreme points of the prediction curve and the time to reach the target temperature.

[0151] When a complex operating condition is identified, historical similar scene data is called up, and a temporal convolutional network is used to predict the long-term temperature change trend, generate a long-term prediction curve, correct the deviation of short-term prediction under complex operating conditions, provide more reliable long-term prediction results, and output the final multi-time scale prediction results, including short-term prediction curve, long-term prediction curve, extreme points, and time to reach the target temperature. Here, complex operating conditions refer to operating conditions under non-steady-state operation scenarios.

[0152] Based on the current operating conditions and equipment status, multiple candidate control strategies are generated, such as adjusting the heating power and cooling valve opening. For each candidate control strategy, a virtual model is used to simulate the impact of each candidate control strategy on the temperature field, and the pre-simulation indicators are recorded, including temperature overshoot, energy consumption, and actuator action count, where the actuator action count is the number of start-stop cycles.

[0153] Taking into account the pre-simulation indicators, the reward value of each candidate control strategy is calculated using a multi-objective optimization algorithm. The candidate control strategy with the highest reward value is selected as the pre-action plan to ensure a balance between temperature accuracy, energy consumption, and equipment lifespan. If multiple candidate control strategies have similar reward values, the candidate control strategy with the fewest actuator actions is selected first to reduce equipment wear and tear. The expression is shown below:

[0154]

[0155] In the formula, For the first The reward value of each candidate control strategy. , , These are the weighting coefficients. , , The first Temperature overshoot, energy consumption, and actuator action count for each candidate control strategy;

[0156] The system calculates the prediction error between the predicted temperature and the actual temperature at each key node in real time, extracts error features including mean, standard deviation, and number of consecutive deviations, and triggers the virtual model calibration process when the prediction error exceeds a preset threshold to ensure timely detection and handling of virtual model prediction deviations; otherwise, no calibration is performed.

[0157] Define the optimization range of thermal property parameters to avoid instability of the virtual model caused by parameter mutations. Use the differential evolution algorithm to iteratively optimize the model parameters with the goal of minimizing the prediction error and output the optimal parameter combination.

[0158] The optimal parameter combination is injected into the virtual model to re-verify the prediction accuracy. Once the verification is successful, the model file is updated to achieve online self-calibration of the model and continuously improve the accuracy and adaptability of the prediction module.

[0159] Specifically, decision engine methods include:

[0160] A working condition-strategy mapping table is constructed to enable the control algorithm corresponding to the working condition label to be directly indexed, ensuring the efficiency and stability of furnace temperature control. When the working condition label is 0, adaptive PID control is used to control the temperature in the steady-state operation scenario, and the PID parameters are dynamically adjusted to make the temperature reach the preset effect. When the working condition label is 1, neural network predictive control is used to reduce temperature overshoot. When the working condition label is 2, fuzzy-feedforward composite control is used to quickly respond to load changes. When the working condition label is 3, graph neural network is used to realize the global optimization allocation of power of each furnace body.

[0161] Obtain the output operating condition labels and equipment health status. Based on the operating condition-strategy mapping table, dynamically call the corresponding control algorithm. For example, in the steady-state operation scenario, call the adaptive PID algorithm, and in the multi-furnace collaborative scenario, activate the graph neural network (GNN) algorithm. Load the initial parameters of the corresponding algorithm, such as the initial value of the PID proportional coefficient and the pre-trained weights of the neural network.

[0162] Based on the characteristics of the current operating conditions, such as quality priority or energy saving priority, the weight coefficients of temperature accuracy, energy consumption, equipment operation frequency and equipment health are dynamically adjusted to construct a multi-objective reward function. For example, in the case of frequent start-stop scenarios, the weight of temperature accuracy is emphasized, and the weight of energy consumption is increased during off-peak hours. At the same time, constraints are defined, including the range of the target temperature deviation threshold and the power change rate limit, to provide optimization objectives for the calculation of control quantities.

[0163] During steady-state operation, the real-time temperature deviation and rate of change are calculated, and the PID parameters are dynamically adjusted through an adaptive algorithm. For example, the proportional coefficient is increased when the deviation is large, and the derivative coefficient is adjusted when the rate of change is fast. The basic control quantities of the heating / cooling equipment are calculated.

[0164] When there are frequent starts and stops or sudden load changes, neural networks or fuzzy-feedforward composite control algorithms are used, combined with temperature trend prediction from the prediction module, to generate control quantities that adapt to nonlinear and time-varying characteristics, such as adjusting the heating power in stages during the start-up phase to avoid overshoot.

[0165] When multiple furnaces work together, the thermal coupling relationship between furnaces is modeled by graph neural network, and the power allocation of each furnace is optimized by considering the power grid constraint. For example, the power of non-critical furnaces is reduced during peak power periods and the preheating power is increased during off-peak power periods to generate a collaborative control vector.

[0166] When the equipment is in an early warning state, such as when the effective value of vibration increases or the cumulative number of actions approaches the rated value, the control quantity is softened, the power change rate is limited or the action delay is increased to reduce equipment wear.

[0167] If the equipment enters a fault state, such as when the health index is below the threshold, the safety control strategy is triggered to forcibly reduce the power to a safe range and send a maintenance work order to avoid damage to the equipment due to excessive control.

[0168] The calculated control quantities are converted into standardized signals that the actuators can recognize, such as 4-20mA current signals and Modbus protocol commands, with additional execution time windows and priority labels. For example, multi-furnace coordination commands are set to high priority to ensure that critical furnaces are executed first.

[0169] Verify whether the control quantity is within the equipment safety threshold, such as heating power not exceeding the rated value and valve opening not lower than the minimum limit. If it exceeds the limit, automatically adjust to the safe range and record a warning.

[0170] In multi-furnace collaborative scenarios, check whether the total power exceeds the grid capacity, dynamically allocate power according to priority, prioritize the power of critical furnaces, and reduce the power of non-critical furnaces proportionally to eliminate power conflicts and ensure grid security.

[0171] The system records control command parameters (such as power command value, execution time, and priority), actual actuator action data (such as actual power and valve opening), and control effects (such as temperature overshoot, energy consumption, and number of actions) in real time. These data are then stored in the log database with timestamps to provide data support for subsequent fault tracing and strategy optimization.

[0172] The system calculates the deviation between the control command and the actual execution, such as power deviation and temperature control deviation, and quantifies the control effect by combining the preset scoring rules. When the score is lower than the preset scoring threshold, a feedback correction signal is generated, such as adjusting PID parameters and optimizing neural network weights, and sent to the optimization module to form a closed loop of control, execution and feedback, continuously improving the performance of the control strategy.

[0173] If feedback correction signals are generated three times in a row, the strategy rollback will be automatically triggered, restoring the effective control algorithm of the previous version.

[0174] Specifically, self-evolutionary methods include:

[0175] Acquire temperature prediction error data, including mean error, standard deviation and maximum deviation, analyze the error distribution of key nodes, identify areas where temperature prediction deviation exceeds the standard, such as the edge of the heating zone or near the cooling channel, and generate an error heat map to intuitively display the prediction accuracy of each area, thus solving the prediction deviation problem caused by aging or material changes in twins.

[0176] Based on real-time error data, the differential evolution algorithm is used to optimize the thermal property parameters of the virtual model. By defining the parameter search range, iterative calculation is performed to minimize the prediction error, update the model parameters, and generate a calibrated virtual model. This replaces the inefficiency of manual calibration, realizes the automatic dynamic update of model parameters, ensures that the digital twin always accurately reflects the thermal characteristics of the physical furnace, and improves prediction accuracy.

[0177] Complete control cycle data is extracted, including state vectors, operating condition labels, control actions, and reward values. The data is cleaned, outliers are removed, and the data is stored according to operating conditions. Then, normalization is performed to form a training dataset, which solves the problems of insufficient data samples and inconsistent quality in reinforcement learning, and ensures the reliability and applicability of the training data.

[0178] Intelligent agent models are trained for single-furnace and multi-furnace scenarios respectively. For single-furnace scenarios, the PPO algorithm is used to optimize control actions with temperature field, load and equipment health as the state space. For multi-furnace scenarios, multi-agent reinforcement learning combined with graph neural network is used to model the cooperative relationship between furnaces, minimize global temperature deviation and total energy consumption, and optimize control actions.

[0179] The newly trained control strategy is validated offline in the virtual model, simulating historical fault conditions and comparing the control effects of the new and old strategies, such as mean error and standard deviation. Once the new strategy performs better, it is deployed to the decision module, and a trial period is set to monitor the performance in real time. If problems occur, it automatically reverts to the old strategy to avoid control performance fluctuations caused by model overfitting or abnormal parameters during strategy iteration. The reliability of the new strategy is ensured through offline validation and trial period mechanisms.

[0180] Temperature accuracy, energy consumption, and equipment operating frequency are defined as optimization objectives. The NSGA-II algorithm is used to generate Pareto optimal solution sets. Each solution corresponds to a set of control parameter combinations, providing diverse parameter combinations for different operating conditions, overcoming the limitations of manual parameter tuning, and automatically adjusting the weights of multiple objectives according to real-time operating conditions. For example, steady-state operation focuses on energy saving, while sudden load changes focus on accuracy. Manual intervention is supported to achieve a balance of optimization objectives under different scenarios and avoid scenario adaptation imbalance with fixed weights.

[0181] When an undefined abnormal operating condition is detected, such as when multiple known operating condition labels are triggered simultaneously, relevant data is automatically collected, abnormal features are identified through the isolated forest algorithm, new operating condition labels are generated and added to the reinforcement learning training set, and the few-shot learning algorithm is started to update the control strategy, thereby improving the adaptability to unknown operating conditions and enhancing the system's adaptability to newly emerging or complex abnormal operating conditions. This avoids control failures caused by incomplete operating condition coverage and improves the system's robustness through self-learning.

[0182] Regularly analyze key performance indicators such as temperature control accuracy, unit product energy consumption, and actuator mean time between failures. Use time series analysis models to predict performance trends. When performance degradation is detected, trigger in-depth calibration or strategy optimization, generate long-term reports and make preventive recommendations to address potential problems such as equipment aging in advance.

[0183] In summary, this invention collects temperature, load, and equipment status data from high-temperature sensors in key parts of the furnace body. This data is then processed through edge computing, spatiotemporal registration, and filtering to generate a standardized state vector. Next, a multi-layer feature recognition system identifies the operating conditions and assesses equipment health, outputting the identification vector. Then, using a digital twin to synchronize data, the temperature field is simulated, trends are predicted, and candidate control strategies are pre-simulated to generate a pre-action plan. Combining the operating conditions and prediction results, the corresponding control algorithm is invoked, multi-objective weights are adjusted to generate control commands, and the actuators are driven after verification. Finally, the data is compared to calibrate the model and iterative strategies, enabling system self-evolution and improving control performance.

[0184] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A furnace body temperature control system, characterized in that, include: Real-time monitoring and acquisition of multi-dimensional data of the physical furnace body, preprocessing to generate state vectors, setting working condition discrimination methods, classifying various working conditions, establishing a health assessment model, and real-time monitoring of actuator status; Construct a virtual model of the furnace body, set up a real-time prediction method, simulate the temperature distribution of the furnace body to predict future temperature trends, determine whether to start the pre-simulation mode, and output the optimal solution. Read the operating condition identification and temperature prediction results, set the decision engine method, generate dynamic control commands, and execute them; By comparing the data of the physical furnace body and the digital twin, a self-evolution method is set up to calibrate and correct the virtual model online. The working condition discrimination method includes: Read the temperature data from the state vector, set a fixed time window, and calculate the rate of temperature change within the fixed time window. and temperature gradient ; Calculate the absolute magnitude of load change by combining real-time load quality and rated load value. and relative rate of change ; Define start / stop events, count the number of device start / stop cycles and power adjustment range, and count the number of start / stop cycles within the fixed time window in real time. Calculate start-stop interval ; Set a threshold for the number of start / stop cycles. and start / stop interval threshold Determine if frequent start-stop conditions occur; if and If a scenario is marked as a frequent start-stop scenario, the operating condition label is 1; otherwise, it is not marked. Set the load change threshold to Temperature response threshold is Determine if the load changes suddenly; and If the load change is sudden, the condition is marked as 2; otherwise, it is not marked. Calculate the correlation coefficient of adjacent furnace temperatures and temperature difference And set relevant thresholds. and temperature difference threshold To identify collaborative working conditions; if and If the conditions for multiple furnace bodies working together are not met, the scenario is marked as a multi-furnace body collaborative scenario, and the operating condition label is 3; otherwise, no label is made. If the conditions for multiple furnace bodies working together are not met, the scenario is judged as a steady-state operation scenario, and the operating condition label is 0. A three-layer health assessment model is constructed based on vibration signals, motor current harmonic distortion rate, and the ratio of cumulative start-stop counts to rated start-stop counts. A weighted summation method is used to generate equipment health indicators, and according to the established health grading rules, the health level of the actuator under the current state and the corresponding response measures are output. The operating condition classification results, dynamic characteristic parameters, and equipment health status are integrated to generate an identification vector, which is then logically verified. The dynamic characteristic parameters include the number of start-stops, start-stop intervals, relative load change rate, temperature change rate, and temperature correlation coefficient in multi-furnace collaborative scenarios.

2. The furnace body temperature control system according to claim 1, characterized in that, The specific steps for generating the state vector include: Sensors were deployed at key locations on the furnace body to form a multi-dimensional monitoring network, and the sensor communication links were tested. Various sensors synchronously collect data at preset frequencies, including temperature arrays, real-time load mass, and time-domain vibration sequences; A unified reference clock is used to calibrate the sensor's time, delay compensation is performed on network transmission data, and synchronized data is output. Adaptive Kalman filtering is used to reduce noise in temperature data, and frequency domain transformation is performed on vibration signals to extract key parameters; The pre-processed temperature, load mass, equipment vibration status, and actuator operating status data are integrated into a unified format state vector, and after validity verification, they are transmitted through an industrial communication protocol.

3. The furnace body temperature control system according to claim 2, characterized in that: In the aforementioned working condition discrimination method, multiple thresholds are set, and an adaptive adjustment mechanism is established for the thresholds; The adaptive adjustment mechanism includes: Collect historical data on equipment operation, extract feature parameter values, and calculate statistics as the basis for threshold initialization; Set initial thresholds for each threshold and save them to the constructed threshold benchmark library; Data is collected in real time using a sliding window, the oldest data is removed and new data is added, and the statistics of each feature parameter within the sliding window are calculated. Calculate the deviation rate between the feature parameters and the initial threshold. And set the deviation threshold. To determine whether to perform adaptive adjustments; when If the initial threshold is adjusted, the adjusted threshold is filtered using an exponential smoothing algorithm; otherwise, the original initial threshold is maintained. Identify overlapping scenarios; if there is a conflict in the threshold adjustment direction in the overlapping scenario, use the Pareto optimization algorithm to generate a compromise threshold and update the threshold of the working condition discrimination method in real time. Verify the reasonableness of the threshold, and choose to use or roll back the threshold based on the verification results.

4. The furnace body temperature control system according to claim 3, characterized in that, The specific steps for constructing a virtual model of the furnace body include: The furnace structure data is acquired using a 3D laser scanner, and point clouds are generated and noise is reduced. The furnace body was divided into multiple finite element mesh elements using finite element software, functional areas were defined, key nodes were marked, and the thermal properties parameters and boundary conditions of the furnace body material were input to construct a three-dimensional geometric model. Collect historical operating condition data and select key data points under typical operating conditions as calibration datasets; The thermophysical parameters were calibrated using a Bayesian optimization algorithm, and a virtual model was output after data verification.

5. The furnace body temperature control system according to claim 4, characterized in that, The real-time prediction method includes: The real-time state vector data is converted into input parameters for the virtual model, and the timestamps are synchronized. The temperature field is calculated by solving the heat conduction equation using the finite element method, and corrections are made for the presence of phase transition processes to generate the temperature field matrix. Based on the temperature field matrix, a prediction curve is generated by extrapolation, and the extreme points of the prediction curve and the time to reach the target temperature are calculated. In non-steady-state operation scenarios, historical data is combined to generate long-term prediction curves using temporal convolutional networks, and prediction results at multiple time scales are output. Candidate control strategies are generated based on operating conditions and equipment status, and pre-simulated using the virtual model, with pre-simulation indicators recorded. Based on a multi-objective optimization algorithm, the reward value of the candidate control strategy is calculated, and the strategy with the largest reward value is selected as the pre-action plan. The prediction error is calculated in real time. When the prediction error exceeds a threshold, the virtual model parameters are optimized and the virtual model is updated using a differential evolution algorithm.

6. The furnace body temperature control system according to claim 5, characterized in that, The decision engine method includes: Construct a working condition-strategy mapping table and establish the correspondence between working condition labels and control algorithms; The corresponding control algorithm is invoked based on the operating condition label and equipment health status, and initial parameters are loaded. Based on the characteristics of the working conditions, the weight coefficients of temperature accuracy, energy consumption, equipment operation frequency, and equipment health are dynamically adjusted to construct a multi-objective reward function and define constraints. Using the corresponding control algorithm, calculate the control quantity for different operating conditions; Depending on the state of the device, the control quantity is softened or made safer. The control quantity is converted into a standardized signal and an execution attribute is attached; Verify whether the control quantity is within the device safety threshold. If it exceeds the device safety threshold, automatically adjust it to the safety threshold and record a warning. In a multi-furnace collaborative scenario, once the total power exceeds the grid capacity, power is dynamically allocated according to priority to handle power conflicts. Record control process data, calculate the deviation between control commands and actual execution, score the control effect, generate feedback correction signals based on the scores, and if correction is triggered three times consecutively, revert to the effective control of the previous version.

7. The furnace body temperature control system according to claim 6, characterized in that, The self-evolution method includes: Acquire temperature prediction error data, analyze the error distribution of key nodes, identify areas where temperature prediction deviation exceeds the standard, and generate an error heat map; The thermal property parameters of the virtual model are optimized using the differential evolution algorithm; Extract complete control cycle data and preprocess it to form a training set; For a single furnace, the PPO algorithm is used to optimize control actions with temperature field, load and equipment health as the state space. For multiple furnaces, multi-agent reinforcement learning combined with graph neural network is used to model the cooperative relationship between furnaces, minimize global temperature deviation and total energy consumption, and optimize control actions. The newly trained control strategy is validated offline in the virtual model, historical fault conditions are simulated, the control effects of the new and old control strategies are compared, and the results are updated in real time. Temperature accuracy, energy consumption, and equipment operating frequency are defined as optimization objectives. The NSGA-II algorithm is used to generate Pareto optimal solution sets, and the weights are dynamically adjusted. When an undefined abnormal operating condition is detected, relevant data is automatically collected, abnormal features are identified through the isolated forest algorithm, new operating condition labels are generated, and control strategies are updated. Regularly analyze key performance indicators, use time series analysis models to predict performance trends, and optimize accordingly.

Images