Big data self-learning evolution method for intelligent burning model of hot blast stove

Abstract

The invention belongs to the technical field of intelligent control of industrial heating equipment, and particularly relates to a big data self-learning evolution method for an intelligent burning furnace model of a hot blast stove. According to the method, multi-source data such as fuel components and temperature fields are collected in real time, and a standardized input data set is constructed after wavelet noise reduction and principal component analysis processing. The model can dynamically optimize the air-fuel ratio, predict the hot air temperature and correct the equipment parameter deviation, and stable combustion is achieved. When effective samples are accumulated to a threshold value, the system automatically triggers parameter iteration, key parameters are reserved and secondary parameters are updated by utilizing transfer learning, encryption gradient aggregation and knowledge sharing among multiple furnaces are realized by virtue of federated learning, and distribution and deployment are performed after a global optimization model is formed. And finally, according to the model output, a closed-loop regulation and control fuel valve, a fan and other execution mechanisms are realized, the temperature and energy consumption are monitored in real time to evaluate the evolution effect, and a continuously optimized intelligent control cycle is formed. According to the method, the control precision and the combustion efficiency are remarkably improved, and energy conservation and consumption reduction are effectively achieved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for industrial heating equipment, specifically a big data self-learning evolution method for an intelligent furnace firing model of a hot blast stove. Background Technology

[0002] As a core heating device in industrial production, the combustion efficiency and control precision of hot blast stoves directly affect production energy consumption, product quality, and equipment lifespan. Traditional hot blast stove control often employs fixed PID control or manual experience-based adjustment modes, which have the following significant drawbacks:

[0003] In industrial environments, factors such as electromagnetic interference, fuel composition fluctuations (such as changes in carbon and hydrogen content in coal gas), and flue gas composition disturbances lead to high noise in the collected data, making it difficult for fixed control parameters to adapt to dynamic operating conditions, and easily causing incomplete combustion or excessive oxygen combustion problems.

[0004] The thermal conductivity of refractory materials in hot blast stoves decreases with increasing operating time. Traditional models do not take this degradation factor into account, which leads to a gradual increase in the deviation of temperature field calculation and a continuous decrease in control accuracy.

[0005] The existing model parameters require manual periodic calibration and cannot be iterated autonomously based on real-time operating data. When the operating conditions (such as heating load and combustion air pressure) change, the model needs to be readjusted, which is time-consuming and affects the continuity of production.

[0006] When multiple hot blast stoves are running in parallel, the data of each stove is stored independently and the control logic is isolated, making it impossible to achieve cross-stove knowledge sharing. As a result, some stoves are in a state of inefficient operation for a long time due to insufficient individual experience.

[0007] When abnormalities such as excessive temperature fluctuations or sudden increases in energy consumption occur, manual investigation of the root cause (such as valve failure or sensor malfunction) is required. The response is delayed and the misjudgment rate is high, which can easily lead to equipment failure or production accidents.

[0008] To address the aforementioned issues, there is an urgent need for an intelligent control method that can achieve data purification, autonomous model evolution, multi-furnace collaborative optimization, and anomaly adaptive handling, in order to overcome the limitations of traditional control and meet the industrial production requirements for high efficiency, low consumption, and stability of hot blast stoves. Summary of the Invention

[0009] The purpose of this invention is to provide a big data self-learning evolution method for intelligent furnace firing models of hot blast stoves, in order to solve the problems of weak anti-interference ability, poor adaptability to equipment aging, lack of self-evolution ability of models, and lack of multi-furnace collaboration in the prior art mentioned in the background.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0011] A big data self-learning evolution method for a smart blast furnace model includes the following steps:

[0012] Step 1, Multi-source data acquisition and feature engineering: Real-time acquisition of fuel composition, furnace temperature field, fluid pressure field and flue gas composition data of hot blast stove; wavelet denoising technology to eliminate industrial electromagnetic interference; extraction of core features through principal component analysis and formation of standardized model input dataset to provide a high-quality data foundation for subsequent self-learning evolution.

[0013] Step 2, intelligent furnace model calculation, including:

[0014] Step 201: Calculate the real-time optimal air-fuel ratio based on fuel composition characteristics and flue gas feedback data to ensure the basic value of combustion efficiency;

[0015] Step 202: Based on multiphysics data, output the hot air temperature for the future preset duration through a temperature advance prediction model;

[0016] Step 203: Based on the equipment running time and thermophysical property data, correct the calculation parameters of the thermal conductivity of the refractory material to offset the impact of equipment aging and deterioration on the accuracy of the model.

[0017] Step 3, Big Data Self-Learning Evolution, includes:

[0018] Step 301: Count the effective sample size of the input dataset in real time. When the effective sample size reaches the set threshold, automatically trigger the model parameter iteration to avoid evolutionary bias caused by a single sample.

[0019] Step 302: Using a transfer learning strategy, the importance weights of each parameter of the model are evaluated first through L1 regularization. The effective parameters with high importance weights are retained, and only the low importance parameters are retrained with new samples to shorten the iteration time and avoid loss of effective knowledge.

[0020] Step 303: Using a multi-furnace federated learning architecture, the local optimization parameters of multiple hot blast stoves are transmitted to the central server through encrypted gradients. After weighted aggregation, a global optimization model is generated and then distributed to each stove to achieve knowledge transfer and improve the model's adaptability to different stove types.

[0021] Step 4, Closed-loop control execution: Based on the air-fuel ratio, temperature prediction results, and aging correction values ​​output in the previous steps, adjust the fuel supply valve opening, combustion fan frequency, and hot air valve switching sequence, and monitor hot air temperature fluctuations and unit fuel consumption in real time. If both temperature fluctuations and energy consumption are lower than the set thresholds, the current evolution is deemed effective, and the current model parameters are stored in the knowledge base; if the target is not met, return to step 3 to re-trigger the iteration until the control objective is met.

[0022] According to the above technical solution, the specific process of calculating the real-time optimal air-fuel ratio in step 201 is as follows:

[0023] First, 100 consecutive sets of fuel component vectors and corresponding flue gas data were collected to construct training sample pairs, providing a data foundation for the subsequent evolutionary optimization of the air-fuel ratio correction coefficient.

[0024] Secondly, the basic air-fuel ratio is calculated by a linear mapping function fitted based on the historical best operating conditions. The linear mapping function takes the fuel component vector as input, and its weights and bias terms serve as the basic parameters for subsequent evolutionary optimization.

[0025] Then, the dynamic correction coefficient is iteratively evolved. That is, the air-fuel ratio correction coefficient is defined as consisting of an initial fixed value, an oxygen concentration correction term, and a carbon monoxide concentration correction term. The oxygen concentration correction term is associated with the deviation between the actual and target oxygen concentration, and the carbon monoxide concentration correction term is associated with the deviation between the actual and target carbon monoxide concentration. Every 100 sets of samples are accumulated, the oxygen concentration correction coefficient and the carbon monoxide concentration correction coefficient are iteratively updated using the gradient descent algorithm with the goal of minimizing the loss of incomplete combustion. When the difference in loss between two iterations is less than 1%, the correction coefficient is determined to have converged, and the current round of evolution is completed.

[0026] The final output and feedback of the final air-fuel ratio is the product of the base air-fuel ratio and the dynamic correction coefficient, which is used as the final control air-fuel ratio. Its calculation formula is:

[0027]

[0028] In the formula, This indicates the final required air-fuel ratio. Indicates the base air-fuel ratio. This represents the air-fuel ratio correction factor; After being applied to control, if the carbon monoxide concentration in the flue gas is consistently ≤500ppm and the oxygen concentration is consistently between 8-10%, the current correction coefficient will be stored in the parameter library as the initial value for the next round of evolution.

[0029] According to the above technical solution, in step 202, the hot air temperature for the future preset duration is output through the temperature advance prediction model as follows:

[0030] A hybrid architecture combining attention mechanism and LSTM is employed, with input consisting of multiphysics time-series data spanning approximately 30-50 time steps. The attention mechanism dynamically focuses on key features significantly impacting temperature prediction. The time-series data for temperature, pressure, and flow are weighted and fused, with the weights satisfying normalization conditions. The absolute error of the temperature prediction is calculated for each set of actual hot air temperatures acquired. The error calculation formula is as follows:

[0031]

[0032] in, This represents the absolute error in temperature prediction. The predicted temperature output by the LSTM. This represents the actual measured temperature value. If the error exceeds the preset error threshold, the fusion weights will be dynamically adjusted to make the weights evolve in the direction of reducing the prediction error.

[0033] Every 50 sets of prediction error data are accumulated, the Adam optimizer is used to update the hidden layer weights and bias terms of the LSTM in batches, with the mean squared error as the optimization target. When the mean squared error is reduced by ≥5%, the network parameter evolution is considered effective; otherwise, it reverts to the parameters of the previous round to avoid evolutionary degradation.

[0034] According to the above technical solution, step 203, correcting the thermal conductivity parameter of the refractory material to offset the impact of equipment deterioration on the model accuracy, specifically includes the following steps:

[0035] An initial decay model for the thermal conductivity of refractory materials is established, using the equipment operating time as a variable. The model describes the change in thermal conductivity with operating time through the initial thermal conductivity and an exponential decay term. The model calculation formula is as follows:

[0036]

[0037] In the formula: Let be the thermal conductivity at a certain time t. The initial thermal conductivity of the material. The term represents the exponential decay, where t is the runtime. The initial attenuation coefficient is used; the initial attenuation model of the thermal conductivity of refractory materials provides a benchmark for heat transfer parameters in the non-aging stage of the equipment.

[0038] During the low-load period of the hot blast stove, the temperature response curve of the refractory material is collected by the thermal pulse method. The actual thermal conductivity is calculated based on the Fourier heat conduction equation. The relative deviation between the theoretical thermal conductivity output by the initial attenuation model and the actual thermal conductivity is calculated. If the deviation exceeds the preset threshold, the attenuation coefficient is dynamically adjusted. That is, when the theoretical value is greater than the actual value, the attenuation coefficient is reduced, and vice versa, to ensure that the model is synchronized with the actual aging state of the equipment.

[0039] Substitute the updated attenuation coefficient into the initial attenuation model to obtain the corrected thermal conductivity calculation value. Substitute the corrected thermal conductivity calculation value into the heat transfer calculation model of the hot blast stove to correct the thermal conductivity parameters in the temperature field calculation, so that the deviation between the calculated temperature and the actual temperature is ≤3℃, and complete the dynamic adaptation of the model to equipment aging.

[0040] According to the above technical solution, the optimization of the evolution threshold for model parameter iteration in step 301 is specifically as follows:

[0041] The 3σ criterion is used to remove outlier data in each feature dimension, retaining the effective data set. At the same time, a typical working condition set is defined, and the proportion of working conditions covered by effective data to the total number of typical working conditions is calculated to ensure the diversity of the evolution sample.

[0042] The initial iteration trigger threshold is set to 500 valid samples, and the operating condition coverage is required to be ≥80%. The threshold is then adjusted based on the change in the total control error of the previous evolution. That is, if the total error decreases, the number of samples required for subsequent triggers is reduced to accelerate the evolution; if the total error increases, the number of samples is increased to ensure data support.

[0043] When the effective sample size reaches the current threshold and the operating condition coverage meets the requirements, the model parameter iteration is triggered immediately; otherwise, data continues to be accumulated to ensure that each evolution is based on high-quality, high-coverage samples.

[0044] According to the above technical solution, in step 302, retaining effective parameters through transfer learning specifically involves:

[0045] The normalized importance weights of each parameter in the model are calculated using L1 regularization. The larger the weight, the more significant the parameter's impact on the control effect.

[0046] If the importance weight of a parameter is ≥0.01 for 3 consecutive rounds, it is marked as a core parameter and only fine-tuned in subsequent evolutions to avoid loss of effective knowledge; if the importance weight of a parameter is <0.005 for 3 consecutive rounds, it is marked as a redundant parameter and deleted in the next evolution to simplify the model structure.

[0047] The core parameters are retained and their values ​​are fixed. Only the low-importance non-core parameters are trained with new samples. The training target is the sum of the errors between the predicted and actual values. This reduces the training time while ensuring the accuracy of the model.

[0048] According to the above technical solution, the global knowledge evolution process of multi-furnace federated learning in step 303 is as follows:

[0049] One central server and N edge nodes are configured. Nodes communicate via the MQTTSN protocol, and data transmission uses Paillier homomorphic encryption. Dynamic addition and removal of nodes is supported to adapt to changes in the furnace group size. The central server distributes initial global model parameters to each node. Nodes train their models using local valid data and calculate the parameter gradient differences. Simultaneously, node data quality weights are introduced, calculated using the following formula:

[0050]

[0051] In the formula: This represents the final weight of the i-th object. Let be the combustion efficiency coefficient of the i-th hot blast stove. , Let be the effective sample size of node i. This represents the sum of the importance and size of all N objects. Nodes with high-quality data and large sample sizes contribute more to the global evolution.

[0052] The central server receives the encrypted gradient difference from each node, decrypts it, and aggregates it according to the data quality weight to generate global optimization parameters, avoiding the operating condition bias of a single node. The global optimization parameters are distributed to each node. If the temperature error of more than 80% of the nodes is reduced by ≥3% or the energy consumption is reduced by ≥1%, the global evolution is deemed effective. Otherwise, the node weights are adjusted and the aggregation is re-aggregated.

[0053] When a new hot blast stove is added, the 128-dimensional equipment feature vector of the new hot blast stove is extracted, the cosine similarity with the feature vector of the existing nodes is calculated, and the model parameters of the three nodes with the highest similarity are selected. Rapid deployment is achieved through small-sample fine-tuning.

[0054] According to the above technical solution, in step 1, wavelet denoising is used to eliminate industrial electromagnetic interference, and principal component analysis is used to extract core features to form a standardized model input dataset. Specifically:

[0055] The db4 wavelet basis and soft thresholding are used for noise reduction. The signal-to-noise ratio of the denoised data is calculated for every 1000 sets of data. If the signal-to-noise ratio is <30dB, the number of wavelet decomposition layers is gradually increased until the signal-to-noise ratio is ≥30dB, to ensure that the noise reduction effect adapts to changes in the noise environment.

[0056] The covariance matrix of the fuel component data matrix is ​​calculated and the eigenvalues ​​are solved. Initially, the top k principal components with a cumulative contribution rate ≥90% are extracted. After each round of model evolution, the cumulative contribution rate of the principal components is recalculated. If the model input error corresponding to the k principal components is >5%, k is increased to ensure the integrity of the feature extraction information and adapt to the changes in model complexity.

[0057] The input data is standardized using Zscore. The mean and standard deviation are recalculated every 500 new data sets to avoid standardization bias caused by changes in data distribution and to provide a stable input basis for model evolution.

[0058] According to the above technical solution, when an anomaly occurs during the evolutionary process, adjustments are made through an anomaly adaptive processing evolutionary mechanism, including the following steps:

[0059] The initial anomaly detection threshold is set based on typical operating conditions. The false positive rate is calculated every 100 anomalies processed. The formula for calculating the false positive rate is:

[0060]

[0061] In the formula, The abnormal false positive rate, The number of misjudged anomalies. This represents the total number of exception handling attempts; if If the threshold is greater than 5%, increase the threshold to reduce false positives; if <2%, reducing the threshold to improve anomaly detection sensitivity and achieve dynamic evolution of the threshold as the operating conditions change;

[0062] Construct a mapping model between anomalies and root causes, train it using a decision tree algorithm, and retrain the decision tree and adjust the tree depth and node splitting threshold every 100 sets of anomaly handling data. If the cross-validation accuracy is ≥90%, update the mapping model; otherwise, retain the original model.

[0063] Establish a self-healing strategy library. Each time a new type of exception is handled, the corresponding handling steps are recorded in the library and the effectiveness of the strategy is marked. When the effectiveness of a strategy is less than 80%, strategy optimization is triggered to achieve continuous evolution of the strategy library.

[0064] Based on the above technical solution, the anomaly-root cause mapping model is constructed as follows:

[0065] When a new anomaly type appears in the hot blast stove, the feature vector of the anomaly is extracted, and the similarity with the existing anomaly features of other nodes in the federated learning system is calculated.

[0066] Select the node with the highest similarity and generate the initial parameters of the new hot blast stove root cause localization model by weighting the similarity. Transfer the model knowledge of similar scenarios to the new hot blast stove to avoid training from scratch.

[0067] The initial model was fine-tuned using 10-20 sets of abnormal data from the new hot blast stove. With the root cause localization accuracy as the optimization objective, the stochastic gradient descent algorithm was used for 100 iterations. When the accuracy was ≥85%, local evolution was completed. The optimized model parameters and corresponding abnormal features were then uploaded to the central server as a supplement to global knowledge, realizing the collaborative evolution of cross-furnace abnormality diagnosis knowledge.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] In this invention, wavelet denoising and principal component analysis effectively eliminate industrial interference and data redundancy; the overall control accuracy is significantly improved compared to traditional PID; based on 3σ sample screening and operating condition coverage control, the high quality of evolutionary samples is ensured; transfer learning shortens the iteration time, federated learning enables knowledge sharing among multiple furnaces, and the model's ability to adapt to different furnace types is improved, eliminating the need for regular manual calibration; through air-fuel ratio optimization, temperature prediction, and multi-furnace collaboration, unit fuel consumption is effectively reduced, combustion efficiency is improved, and annual energy costs are significantly reduced. Attached Figure Description

[0070] Figure 1This is a schematic diagram of the overall process of the self-learning evolution method of the present invention;

[0071] Figure 2 This is a flowchart illustrating the calculation process of the intelligent furnace model using the self-learning evolution method of this invention.

[0072] Figure 3 This is a flowchart of the big data self-learning evolution method of the present invention. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] Example 1

[0075] like Figure 1 As shown, this invention comprises four core steps: multi-source data acquisition and feature engineering, intelligent furnace model calculation, big data self-learning evolution, and closed-loop control execution. It also incorporates an anomaly adaptive handling mechanism, as detailed below:

[0076] Step 1, Multi-source Data Acquisition and Feature Engineering: The core objective of this step is to acquire full-condition data of the hot blast stove and perform purification, dimensionality reduction, and standardization to provide a high-quality input foundation for model evolution. Specifically, this includes:

[0077] Real-time acquisition of multi-source data: Four types of core data are collected through online sensors and analyzers: fuel composition data, such as the mass fraction of carbon (C), hydrogen (H) and oxygen (O) in coal gas, with a sampling frequency of 1 time / minute.

[0078] Furnace body temperature field data: The temperature of 68 key areas, including the top of the furnace, side walls, and air outlet, was collected using an infrared thermometer at a sampling frequency of 1 time / 10 seconds.

[0079] Fluid pressure field data: Combustion air inlet pressure, gas inlet pressure, and flue gas outlet pressure are collected by pressure transmitters at a sampling frequency of 1 time / 5 seconds.

[0080] Flue gas composition data: Oxygen (O2) and carbon monoxide (CO) in flue gas were collected using a gas analyzer at a sampling frequency of 1 time / 30 seconds.

[0081] Wavelet noise reduction: To address data noise caused by industrial electromagnetic interference, a multi-level decomposition using the dB4 wavelet basis is employed, and noise is suppressed through a soft thresholding method, as detailed below:

[0082] Soft threshold calculation: Threshold , where σ is the noise standard deviation (estimated through data smoothing segments). This represents the data length.

[0083] Noise reduction effect verification: For every 1000 sets of data accumulated, the signal-to-noise ratio (SNR) of the noise-reduced data is calculated using the following formula:

[0084]

[0085] in This is the original data. The data is after noise reduction. If SNR < 30dB, the number of wavelet decomposition layers (initially set to 3 layers) is increased by 1 layer until SNR ≥ 30dB, ensuring that data noise is effectively suppressed.

[0086] Principal Component Analysis (PCA) Dimensionality Reduction: Addressing the redundancy in multi-dimensional data, PCA extracts core features.

[0087] Covariance matrix calculation: for fuel component data vectors (m1 is the number of samples, n is the number of groups), calculate the covariance matrix. ,in For sample vectors, This is the sample mean vector.

[0088] Core feature extraction: Select the top k principal components with a cumulative contribution rate of ≥90% as core features; after each round of model evolution, if the model input error corresponding to the k principal components is >5%, then increase k (up to n) to ensure the integrity of feature information.

[0089] Data standardization: The core feature data is normalized using the Zscore standardization formula, which is as follows:

[0090]

[0091] Where μ is the data mean and σ is the noise standard deviation; every 500 sets of new data are accumulated, μ and σ are recalculated to avoid standardization bias caused by changes in data distribution, and finally a standardized model input dataset is formed.

[0092] Step 2, Intelligent furnace model calculation (e.g.) Figure 2 (As shown): This step constructs a multi-objective optimization model based on standardized data. Through air-fuel ratio optimization, temperature advance prediction, and equipment aging correction, it provides accurate parameters for control execution. Specifically, it includes three sub-steps:

[0093] Step 201, Real-time Optimal Air-Fuel Ratio Calculation (Ensuring Baseline Combustion Efficiency): The air-fuel ratio is a core parameter affecting combustion efficiency. This step achieves the optimal value output through basic calculations and dynamic corrections.

[0094] Model input layer evolution preparation: Collect 200 consecutive sets of fuel component vectors (including C, H, and O content) and corresponding flue gas data (O2 and CO concentration) to construct training sample pairs for evolutionary optimization of correction coefficients;

[0095] Basic air-fuel ratio calculation: The initial baseline value is obtained through a linear mapping function, the formula is as follows:

[0096]

[0097] Where X is the fuel component data vector, which contains the values ​​of each component in the fuel (such as carbon content, hydrogen content, oxygen content, etc.), a is the component weight vector (such as C weight 0.6, H weight 0.3, O weight 0.1), and b is the bias term (initially set to 0.5). a and b are obtained by fitting the historical best operating condition data using the least squares method.

[0098] Iterative evolution of dynamic correction coefficients: introducing correction coefficients The base air-fuel ratio is dynamically adjusted to ensure complete combustion, as detailed below:

[0099] Correction coefficient formula:

[0100]

[0101] in: The initial fixed value is 1.05~1.15, which does not participate in evolution. The target oxygen concentration is 8-10%. This represents the actual oxygen concentration. The target CO concentration is ≤500ppm. This represents the actual CO concentration. This is the oxygen concentration correction factor. , which is the carbon monoxide concentration correction factor (both are evolutionary variables);

[0102] Evolutionary Objective and Loss Function: With the objective of minimizing the loss from incomplete combustion, the loss function L is defined as follows:

[0103]

[0104] In the formula, For combustion loss, For the target oxygen concentration, This represents the actual oxygen concentration. For the target CO concentration, This represents the actual CO concentration.

[0105] For every 100 sets of samples accumulated, k1 and k2 are updated using the gradient descent algorithm. The update formula is as follows:

[0106]

[0107]

[0108] in, This represents the parameter value after iterative updates of the oxygen concentration correction coefficient. This represents the parameter value before the oxygen concentration correction coefficient was iteratively updated; This represents the parameter value after iterative updates of the carbon monoxide concentration correction coefficient. This represents the parameter value before the iterative update of the carbon monoxide concentration correction coefficient; The learning rate is used; when the difference in L between two iterations is less than 1%, a decision is made. , Convergence completes this round of evolution.

[0109] Final air-fuel ratio output: optimal air-fuel ratio ;Will After being applied to control, if Stable ≤500ppm and If it stabilizes at 8-10%, then the current... , Store it in the parameter library as the initial value for the next round of evolution.

[0110] Step 202, Hot Air Temperature Advance Prediction (Adapting to Changes in Operating Conditions in Advance). A hybrid architecture combining attention mechanisms and LSTM is used to predict the hot air temperature 10-30 minutes in advance (the preset duration can be adjusted as needed), as detailed below:

[0111] Input data and architecture design: Input multiphysics time-series data (including temperature, pressure, and flow) for approximately 30-50 time steps. An attention mechanism is used to dynamically focus on key features (such as pressure data during load abrupt changes), and LSTM is used to capture temporal dependencies.

[0112] Feature fusion: Multiphysics feature fusion is achieved through weighted summation, and the formula is as follows:

[0113]

[0114] Where V is the fusion feature vector; , , These are time series sequences for temperature, pressure, and flow rate, respectively; ω1, ω2, and ω3 are the feature weights for temperature, pressure, and flow rate (satisfying ω1+ω2+ω3=1).

[0115] Weighting evolution update: For each set of actual hot air temperatures obtained... Calculate and predict temperature Error with actual temperature If E > the preset error threshold (e.g., ±3℃), then the weights are updated according to the following formula:

[0116]

[0117]

[0118] in The denominator is set to the learning rate, ensuring that the sum of the weights is always 1. This process causes the weights to evolve in a direction that reduces prediction error.

[0119] LSTM parameter batch evolution: Every 50 sets of prediction error data are accumulated, the hidden layer weights and bias terms of the LSTM are updated in batches, using the Adam optimizer, with the mean squared error (MSE) as the metric. The objective function is m2 (where m is the number of samples). When the MSE decreases by ≥5%, the network parameter evolution is considered effective; otherwise, it reverts to the parameters of the previous round to avoid evolutionary degradation.

[0120] Step 203, Correction of Refractory Material Thermal Conductivity (To Counteract Equipment Deterioration). To address the decrease in thermal conductivity caused by refractory material aging, dynamic correction is achieved through initial model and actual calibration.

[0121] Initial decay model construction: The initial thermal conductivity λ(t) is modeled according to an exponential decay law, and the formula is:

[0122]

[0123] in Let be the thermal conductivity at a certain time t. The initial thermal conductivity of the material. The term represents the exponential decay, where t is the runtime. The initial attenuation coefficient is used to establish the aging baseline.

[0124] Actual parameter calibration: During the low-load period of the hot blast stove (load <30%, to avoid production interference), the actual thermal conductivity is measured by the thermal pulse method: a 50ms, 100W rectangular thermal pulse is emitted onto the refractory material surface, and the temperature response curve from 0 to 10 seconds after the pulse is collected. The actual thermal conductivity is calculated based on the Fourier heat conduction equation. .

[0125] Attenuation coefficient evolution: calculating the deviation rate between theoretical and actual values. ;like If a preset deviation threshold (e.g., 5%) is set, the attenuation coefficient will be updated using the following formula:

[0126]

[0127] when hour, , To make the attenuation model fit the actual aging rate, increase k; otherwise, increase k to ensure that the model is synchronized with the equipment status.

[0128] Heat transfer model adaptation: The updated model will be adapted to the new model. Substitute the heat transfer calculation model of the hot blast stove, correct the heat conduction parameters in the temperature field calculation, and make the deviation between the calculated temperature and the actual temperature ≤3℃, thus completing the dynamic adaptation of equipment aging.

[0129] Step 3, Big Data Self-Learning Evolution (e.g.) Figure 3 (As shown). This step achieves autonomous iteration of model parameters and knowledge sharing across multiple furnaces through sample selection, transfer learning, and federated learning, ensuring continuous optimization of the model as operating conditions change. Specifically, it includes three sub-steps:

[0130] Step 301, Model Iteration Trigger Condition Optimization (Ensuring Sample Quality). To avoid evolutionary bias caused by a single or insufficient sample, the iteration trigger timing is determined through anomaly removal and operational condition coverage:

[0131] Outlier removal is performed using the 3σ criterion:

[0132] Operating condition coverage calculation: Define a typical operating condition set S (including 10-15 operating conditions such as high / medium / low load and different fuel compositions), and calculate the number of operating conditions S effectively covered by data. cover Then the operating condition coverage (|S| represents the total number of operating conditions for S), and operating condition coverage ensures the diversity of evolutionary samples.

[0133] Dynamic trigger threshold: The initial trigger threshold is set to... Group (effective sample size) (Operating condition coverage); As the model evolves, the trigger threshold is dynamically adjusted according to the following formula:

[0134]

[0135] in The adjustment factor is 0.05. , Total control error ( , This represents the average temperature error. (This refers to the average energy consumption error).

[0136] When the effective sample size is ≥ and If the condition is met, parameter iteration is triggered immediately; otherwise, data accumulation continues.

[0137] Step 302: Transfer learning preserves effective knowledge (shortening iteration time). Parameter importance is evaluated using L1 regularization, and only low-importance parameters are trained to avoid loss of effective knowledge and redundant computation.

[0138] Parameter importance weight calculation: The normalized importance weights of each model parameter are calculated using L1 regularization, and the formula is as follows:

[0139]

[0140] In the formula, The larger the value, the more significant the impact on the control effect, and it should be retained first.

[0141] Parameter classification label: recalculated after each iteration. :like If the value is ≥0.01 for three consecutive rounds, it is marked as a core parameter. In subsequent evolution, it is only fine-tuned (adjustment range ≤5%) to avoid the loss of core knowledge.

[0142] like If the value is less than 0.005 for three consecutive rounds, it is marked as a redundant parameter and will be deleted in the next evolutionary round to simplify the model structure.

[0143] Effective parameter preservation and new parameter training: Preservation Parameters ≥0.008 are only applicable to Parameters with a value < 0.008 are trained using new data, with the training objective being the total prediction error. :

[0144]

[0145] Stochastic gradient descent algorithm is used for optimization, with 500 training epochs and a learning rate of [missing information]. This significantly reduces iteration time (more than 60% less than full-parameter training).

[0146] Step 303: Multi-furnace Federated Learning Enables Knowledge Sharing (Improving Equipment Adaptability). An architecture consisting of a central server and edge nodes is constructed to achieve collaborative knowledge evolution across multiple hot blast furnaces, while simultaneously protecting data privacy.

[0147] Architecture design: Set up 1 central server and N edge nodes (each node corresponds to 1 hot air furnace). The nodes communicate with each other via the MQTTSN protocol (suitable for low power consumption in industrial scenarios). Data transmission adopts Paillier homomorphic encryption (to ensure that the original data is not leaked during parameter transmission).

[0148] The central server issues the initial global model parameters. To each edge node; node i uses locally valid data Train the model to obtain local model parameters Calculate the gradient difference of the parameters ( The gradient difference of the parameters reflects the local evolutionary contribution of node i; a positive value indicates that the parameters need to be increased, and a negative value indicates that they need to be decreased.

[0149] Introducing node data quality weights Nodes with high-quality data and large sample sizes contribute more to global evolution, as shown in the formula:

[0150]

[0151] in This represents the final weight of the i-th object. Let i be the operating condition coverage. This represents the effective sample size of node i. The central server receives encrypted data from each node. and After decryption, the global optimization parameters are aggregated according to the following formula:

[0152]

[0153] global optimization parameters Distribute to each node for testing. Control effectiveness on local data: If the temperature error of more than 80% of nodes decreases by ≥3% or energy consumption decreases by ≥1%, then global evolution is considered effective; otherwise, adjustments are made. Weight (e.g., increase) (High node weight), re-aggregate.

[0154] Rapid Deployment of New Hot Blast Stove: When a new hot blast stove is added to the system, its equipment characteristics (such as stove type, refractory material type, and rated load) are extracted to construct a 128-dimensional feature vector. Calculate the feature vectors of each edge node. Cosine similarity:

[0155]

[0156] Select The model parameters of the three largest nodes are fine-tuned using 10-20 sets of local data from the new hot air furnace, enabling rapid deployment (shortening the time by more than 90% compared to training from scratch).

[0157] Step 4: Closed-loop control execution and adaptive exception handling.

[0158] Step 401, Closed-loop control execution: Based on the optimal air-fuel ratio output by the model ( Temperature prediction results ), aging correction value ( ), and perform the following control actions:

[0159] Adjust the fuel supply valve opening (according to) (The opening adjustment accuracy is ±0.5%).

[0160] Adjust the frequency of the combustion fan (the frequency adjustment accuracy is ±0.1Hz, depending on the pressure field requirements).

[0161] Adjust the hot air valve switching sequence (according to) (Timing adjustment accuracy is ±5 seconds).

[0162] Real-time monitoring of hot air temperature fluctuations ( ) and unit fuel consumption ( If ΔT ≤ preset threshold (e.g., ±2℃) and E fule If the energy consumption threshold is met, the current evolution is considered valid, and the current model parameters are stored in the knowledge base; if the threshold is not met, return to step 3 to re-trigger the iteration until the control objective is met.

[0163] Step 402, Anomaly Adaptive Handling. When anomalies such as excessive temperature fluctuations or sudden increases in energy consumption occur, autonomous adjustment is achieved through the following mechanisms:

[0164] Anomaly detection threshold evolution: Initial temperature anomaly threshold =8℃, energy consumption anomaly threshold =10%; Calculate the false positive rate for every 100 exceptions processed. ;like >5%, then according to Increase the threshold to reduce false positives; if If the threshold is less than 2%, then decrease the threshold ( ), to improve detection sensitivity.

[0165] Root cause localization model iteration: Constructing a mapping function between anomalies and root causes (such as valve jamming, sensor drift, and fuel composition mutations). ,in For temperature error, The energy consumption error is represented by R as the root cause. A decision tree algorithm is used for training, and the splitting criterion is the Gini coefficient. (D is the sample set, The percentage of samples representing the r-th root cause); every 100 sets of anomaly handling data accumulated (including , (R, processing effect), retrain the decision tree, adjust the tree depth and node splitting threshold; if the cross-validation accuracy... If the mapping function is correct, then update the mapping function F; otherwise, retain the original model.

[0166] Evolution of the self-healing strategy library: Establish a self-healing strategy library. Every time a new type of anomaly is handled (such as a sudden increase in CO and valve leakage), the corresponding handling steps (such as closing the fuel valve, repairing the valve, restarting and debugging) are entered into the library, and the effectiveness of the strategy is marked (effectiveness = number of successful handling / number of applications). When the effectiveness of a certain strategy is less than 80%, strategy optimization is triggered to achieve continuous improvement of the strategy library.

[0167] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0168] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A big data self-learning evolution method for a smart blast furnace model, characterized in that: Comprising the following steps: Step 1, multi-source data acquisition and feature engineering: real-time acquisition of hot blast stove fuel components, furnace temperature field, fluid pressure field and flue gas composition data, using wavelet denoising technology to eliminate industrial electromagnetic interference, extracting core features through principal component analysis and forming standardized model input data set, providing high-quality data basis for subsequent self-learning evolution; Step 2, intelligent stove model calculation, including: Step 201, based on fuel component characteristics and flue gas feedback data, calculate the real-time optimal air-fuel ratio to ensure the basic value of combustion efficiency; Step 202, based on multi-physical field data, output the hot blast temperature in the future preset time through the temperature advance prediction model; Step 203, based on equipment running time and thermal physical property data, correct the thermal conductivity coefficient calculation parameters of refractory materials to offset the influence of equipment aging and degradation on model accuracy; Step 3, big data self-learning evolution, including: Step 301, real-time statistics of the effective sample size of the input data set, when the effective sample size reaches the set threshold, automatically trigger model parameter iteration to avoid evolution deviation caused by single sample; Step 302, use transfer learning strategy, first evaluate the importance weight of each parameter of the model through L1 regularization, retain the effective parameters with high importance weight, only retrain the low importance parameters with new samples, shorten the iteration time, and avoid losing effective knowledge; Step 303, use multi-furnace federated learning architecture, transfer the local optimization parameters of multiple hot blast stoves to the central server through encrypted gradient, aggregate the global optimization model through weighted aggregation, and then distribute it to each furnace to realize knowledge transfer and improve the adaptability of the model to different furnace types; Step 4, closed-loop control execution: adjust the fuel supply valve opening, combustion air fan frequency and hot blast valve switching time according to the air-fuel ratio, temperature prediction results and aging correction values output by the previous steps, real-time monitor the hot blast temperature fluctuation and unit fuel consumption, if the temperature fluctuation and energy consumption are lower than the set threshold, it is determined that the current evolution is effective, and the current model parameters are stored in the knowledge base; if not, return to step 3 to trigger iteration again until the control target is met.

2. The big data self-learning evolution method of the hot blast stove intelligent burning model according to claim 1, characterized in that: In step 201, the specific process of calculating the real-time optimal air-fuel ratio is as follows: First, collect 100 groups of continuous fuel component vectors and corresponding flue gas data to construct training samples for subsequent evolution optimization of air-fuel ratio correction coefficients; Second, calculate the basic air-fuel ratio through the linear mapping function based on the historical optimal working condition fitting, and the weight and bias term of the fuel component vector as the basis for subsequent evolution optimization; Then, iterate and evolve the dynamic correction coefficient, that is, define the air-fuel ratio correction coefficient composed of the initial fixed value, oxygen concentration correction term and carbon monoxide concentration correction term, wherein the oxygen concentration correction term is related to the deviation between the actual and target oxygen concentration, and the carbon monoxide concentration correction term is related to the deviation between the actual and target carbon monoxide concentration. Every 100 samples are accumulated to update the oxygen concentration correction coefficient and the carbon monoxide concentration correction coefficient through gradient descent algorithm iteration, and when the loss difference of two iterations is less than 1%, the correction coefficient is considered to be converged, and the current evolution is completed. The final output and feedback of the final air-fuel ratio is the product of the basic air-fuel ratio and the dynamic correction coefficient as the final control air-fuel ratio, and the calculation formula is: In the formula, represents the final required air-fuel ratio, represents the base air-fuel ratio, represents the air-fuel ratio correction coefficient; After application of the control, if the carbon monoxide concentration in the flue gas is stable at ≤500 ppm and the oxygen concentration is stable at 8-10%, the current correction coefficient is stored in the parameter library as the initial value for the next round of evolution.

3. The big data self-learning evolution method of the hot blast stove intelligent burning model according to claim 2, characterized in that: In step 202, the hot blast temperature in the future preset time is output by the temperature lead prediction model, specifically: The attention mechanism and LSTM hybrid architecture are adopted, the input is the multi-physical field time series data of the last 30-50 time steps, the key features that significantly affect the temperature prediction are dynamically focused through the attention mechanism; the temperature, pressure and flow time series are weighted and fused, the fusion weight satisfies the normalization condition, the temperature prediction absolute error is calculated every time a group of actual hot blast temperature is obtained, and the error calculation formula is: wherein, represents the absolute error of temperature prediction, is the predicted temperature output by the LSTM, represents the actually measured temperature value, if is greater than a preset error threshold, the fusion weight is dynamically adjusted to evolve in the direction of reducing the prediction error. Every 50 groups of prediction error data are accumulated, the Adam optimizer is used to update the hidden layer weight and bias of LSTM in batches, the mean square error is taken as the optimization target, when the mean square error is reduced by ≥5%, it is determined that the network parameter evolution is effective, otherwise it is rolled back to the last round of parameters to avoid evolution degradation.

4. The big data self-learning evolution method of the hot blast stove intelligent burning model according to claim 3, characterized in that: For step 203, the thermal conductivity coefficient of refractory material is corrected to offset the influence of equipment degradation on model accuracy, which includes the following steps: An initial attenuation model of the thermal conductivity coefficient of refractory material is established, which takes the running time of the equipment as the variable, and describes the change rule of the thermal conductivity coefficient with the running time through the initial thermal conductivity coefficient and the exponential attenuation term, and the model calculation formula is: wherein: k(t) is the thermal conductivity at a time t, k0 is the initial thermal conductivity of the material, exp is the exponential decay term, and t is the time of operation, k0 is the initial thermal conductivity of the material; the initial decay coefficient provides a baseline for the heat transfer parameters for the unaged stage of the equipment; During the low load period of the hot blast stove, the temperature response curve of the refractory material is collected by the heat pulse method, the actual thermal conductivity coefficient is calculated by inverse calculation based on the Fourier heat conduction equation, the relative deviation between the theoretical thermal conductivity coefficient output by the initial attenuation model and the actual thermal conductivity coefficient is calculated, and if the deviation exceeds the preset threshold, the attenuation coefficient is dynamically adjusted, that is, when the theoretical value is greater than the actual value, the attenuation coefficient is reduced, and vice versa, to ensure that the model is synchronized with the actual aging state of the equipment; The updated attenuation coefficient is substituted into the initial attenuation model to obtain the corrected thermal conductivity calculation value, and the corrected thermal conductivity calculation value is substituted into the hot blast stove heat transfer calculation model to correct the thermal conductivity term parameter in the temperature field calculation, so that the deviation between the calculated temperature and the actual temperature is ≤3℃, and the dynamic adaptation of the model to the equipment aging is completed.

5. The big data self-learning evolution method of the hot blast stove intelligent burning model according to claim 4, characterized in that: In step 301, the evolution threshold optimization of model parameter iteration is as follows: Abnormal data in each feature dimension is removed by 3σ criterion, and effective data set is reserved, and typical working condition set is defined, the proportion of working condition covered by effective data in total number of typical working conditions is calculated to ensure diversity of evolution samples; The initial iteration trigger threshold is set to 500 groups of effective samples, and the working condition coverage degree is required to be ≥80%, and the threshold is adjusted in combination with the total control error change of the last round of evolution, that is, if the total error is reduced, the sample amount required for subsequent triggering is reduced to speed up evolution; If the total error increases, increase the sample size to ensure data support; When the effective sample size reaches the current threshold and the working condition coverage degree meets the requirements, the model parameter iteration is triggered immediately, otherwise the data is accumulated to ensure that each evolution is based on high-quality and high-coverage samples.

6. The big data self-learning evolution method of the hot blast stove intelligent burning model according to claim 5, characterized in that: In step 302, the effective parameters are reserved by transfer learning, specifically: The normalized importance weight of each parameter of the model is calculated by L1 regularization, and the greater the weight, the more significant the parameter's impact on the control effect; If the importance weight of a parameter is ≥0.01 for 3 consecutive rounds, it is marked as a core parameter, and only fine-tuning is performed in subsequent evolution to avoid effective knowledge loss; if the importance weight of a parameter is <0.005 for 3 consecutive rounds, it is marked as a redundant parameter, and is deleted in the next round of evolution to simplify the model structure; The core parameters are retained and fixed, and only the low-importance non-core parameters are trained with new samples to predict the error sum of the actual value, which shortens the training time while ensuring the accuracy of the model.

7. The big data self-learning evolution method of the hot blast stove intelligent burning model according to claim 6, characterized in that: In step 303, the global knowledge evolution process of multi-furnace federated learning is as follows: Set up 1 central server and N edge nodes, communicate between nodes through MQTTSN protocol, data transmission uses Paillier homomorphic encryption, support dynamic increase and decrease of nodes to adapt to the change of furnace group size; the central server issues the initial global model parameters to each node, the node trains the model with local effective data and calculates the parameter gradient difference, and introduces the node data quality weight, the weight calculation formula is: In the formula: represents the maximum weight of the ith object, is the combustion efficiency coefficient of the ith hot blast stove, , is the effective sample size of node i, represents the sum of the importance and size of all N objects, The node with high data quality and large sample size contributes more to global evolution. The central server receives the encrypted gradient difference of each node, decrypts and aggregates the global optimization parameters according to the data quality weight to avoid the bias of single node working condition; Distribute the global optimization parameters to each node, if the temperature error of more than 80% of the nodes is reduced by ≥3% or the energy consumption is reduced by ≥1%, determine that the global evolution is effective, otherwise adjust the node weight and re-aggregate; When a new hot blast stove is added, extract the 128-dimensional equipment feature vector of the new hot blast stove, calculate the cosine similarity with the existing node feature vector, select the top 3 node model parameters with the highest similarity, and realize rapid deployment through small sample fine-tuning.

8. The big data self-learning evolution method of the hot blast stove intelligent burning model according to claim 1, characterized in that: In step 1, wavelet denoising is used to eliminate industrial electromagnetic interference, and principal component analysis is used to extract core features to form a standardized model input data set, which is as follows: Use db4 wavelet basis and soft threshold denoising, calculate the signal-to-noise ratio of the denoised data every 1000 groups of data, if the signal-to-noise ratio is <30dB, gradually increase the wavelet decomposition level until the signal-to-noise ratio is ≥30dB, and ensure that the denoising effect adapts to the change of noise environment; Calculate the covariance matrix of the fuel component data matrix and solve the eigenvalues, initially extract the first k principal components with a cumulative contribution rate of ≥90%, and recalculate the cumulative contribution rate of the principal components after each round of model evolution, if the model input error corresponding to the k principal components is >5%, increase k to ensure the information integrity of feature extraction adapts to the change of model complexity; Use Zscore standardization to process input data, recalculate the mean and standard deviation every 500 new data to avoid standardization deviation caused by data distribution change, and provide stable input basis for model evolution.

9. The hot blast stove intelligent stove model big data self-learning evolution method according to claim 1, characterized in that: When an abnormality occurs in the evolution process, the evolution mechanism of abnormal self-adaptive processing is used for adjustment, including the following steps: The initial abnormality judgment threshold is set according to the typical working condition, and the misjudgment rate is calculated every 100 times of abnormality processing, the misjudgment rate calculation formula is: The central server receives the encrypted gradient difference of each node, decrypts and aggregates the global optimization parameters according to the data quality weight to avoid the bias of single node working condition; In the formula, is the abnormal misjudgment rate, is the number of misjudged abnormalities, is the total number of abnormal treatments; if > 5%, increase the threshold value to reduce misjudgment; if < 2%, reduce the threshold value to improve the sensitivity of abnormal detection, and realize the dynamic evolution of the threshold value with the change of the working condition; The mapping model of the abnormality and the root cause is constructed, a decision tree algorithm is trained, 100 groups of abnormality processing data are accumulated, the decision tree is retrained and the tree depth and node split threshold are adjusted, if the cross-validation accuracy is greater than or equal to 90%, the mapping model is updated, otherwise the original model is reserved; A self-healing strategy library is established, each time a new type of abnormality is processed, the corresponding processing steps are recorded in the library and the strategy validity is marked, when the validity of a strategy is less than 80%, the strategy optimization is triggered, and the continuous evolution of the strategy library is realized.

10. The big data self-learning evolution method of the hot blast stove intelligent burning model according to claim 9, characterized in that: The mapping model of the abnormality and the root cause is constructed, a decision tree algorithm is trained, 100 groups of abnormality processing data are accumulated, the decision tree is retrained and the tree depth and node split threshold are adjusted, if the cross-validation accuracy is greater than or equal to 90%, the mapping model is updated, otherwise the original model is reserved; A self-healing strategy library is established, each time a new type of abnormality is processed, the corresponding processing steps are recorded in the library and the strategy validity is marked, when the validity of a strategy is less than 80%, the strategy optimization is triggered, and the continuous evolution of the strategy library is realized. The mapping model of the abnormality and the root cause is constructed, a decision tree algorithm is trained, 100 groups of abnormality processing data are accumulated, the decision tree is retrained and the tree depth and node split threshold are adjusted, if the cross-validation accuracy is greater than or equal to 90%, the mapping model is updated, otherwise the original model is reserved; A self-healing strategy library is established, each time a new type of abnormality is processed, the corresponding processing steps are recorded in the library and the strategy validity is marked, when the validity of a strategy is less than 80%, the strategy optimization is triggered, and the continuous evolution of the strategy library is realized.

Images