Intelligent blast furnace condition trend early warning system and method

Abstract

The application discloses an intelligent blast furnace condition trend early warning system and method, and relates to the field of blast furnace condition monitoring.The system comprises a data acquisition module, a dynamic detection module, a quality correlation module and a feedback control module.The data acquisition module generates original characteristic signals correlated with time and space; the preprocessing module receives the original characteristic signals and filters out noise interference through a Gaussian process regression algorithm; the prediction module generates correlation characteristics of coal gas flow distribution and material surface morphology, and then captures long-period trend components in the standardized data stream through a time convolution network; the feedback module receives blast furnace condition evolution probability signals and generates virtual blast furnace state data based on a physical field simulation model of gas-solid two-phase flow and heat transfer coupling; and the early warning module generates graded early warning instructions according to the degree of parameter calibration signals crossing a preset threshold.The application can solve the problem of early warning lag caused by the dependence of traditional blast furnace condition early warning methods on single parameters, static models and artificial experience.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and fault early warning technology for blast furnace ironmaking processes, specifically to an intelligent blast furnace condition trend early warning system and method. Background Technology

[0002] As a core piece of equipment in steel production, the blast furnace involves a highly coupled process of gas-solid heat transfer, chemical reaction, and material movement. The stability of the furnace condition directly affects production efficiency and safety. Traditional monitoring methods mainly rely on single-parameter threshold alarms such as temperature and pressure. However, blast furnaces are highly enclosed and operate under complex conditions, making it difficult for a single parameter to comprehensively reflect the internal state. Although machine learning models (such as LSTM) have been introduced for trend prediction in recent years, they still face two bottlenecks due to limitations such as single data dimensions and static models: First, at the data level, only traditional process parameters are collected, ignoring multi-physical field information such as vibration, acoustics, and thermal imaging, resulting in incomplete feature representation. Second, at the model level, traditional statistical models or single deep learning architectures cannot simultaneously characterize the spatial correlation and temporal dynamics between blast furnace parameters and lack online self-calibration capabilities.

[0003] For example, a steel plant used a BP neural network to predict furnace temperature, but because it did not consider the spatial coupling relationship between the gas flow and the material surface morphology, the false alarm rate was consistently higher than 15%. Furthermore, existing digital twin technologies are mostly used for equipment simulation and are not deeply integrated with real-time early warning systems, resulting in model parameter updates lagging behind actual operating conditions. With the intelligent transformation of the steel industry, multi-source heterogeneous data fusion and dynamic model optimization have become key technological breakthroughs. However, existing solutions have not yet formed systematic solutions in areas such as data processing (e.g., noise filtering and feature extraction), model architecture (e.g., spatiotemporal feature co-modeling), and virtual-real interaction (e.g., twin-driven closed-loop optimization), making it difficult for early warning systems to meet the requirements of long-term stable operation of blast furnaces. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide an intelligent blast furnace condition trend early warning system and method to solve the problem of delayed early warning caused by the reliance on single parameters, static models and human experience in traditional blast furnace condition early warning methods. This invention collects multi-dimensional data through vibration, ultrasonic and infrared sensors. After noise filtering and standardization, it uses graph convolutional networks to extract spatial correlation features between parameters, combines them with temporal convolutional networks to capture long-term trends, and dynamically weights key events through an attention mechanism. A digital twin module compares virtual simulation data with actual operating conditions in real time, generates parameter calibration signals to optimize the prediction model, and an early warning module outputs operational suggestions in stages according to the degree of deviation between the calibration signal and the threshold, forming a closed loop of "perception-prediction-calibration-decision," ultimately solving the problem of poor adaptability of traditional methods to complex operating conditions.

[0005] This invention provides an intelligent blast furnace condition trend early warning system, comprising:

[0006] The data acquisition module collects real-time parameters of the blast furnace and generates spatiotemporally correlated raw characteristic signals.

[0007] The preprocessing module receives the original feature signal and filters out noise interference through a Gaussian process regression algorithm to form a denoised and standardized data stream.

[0008] The prediction module receives a standardized data stream, extracts the spatial topological relationship between blast furnace parameters through a graph convolutional network, generates the correlation features between gas flow distribution and material surface morphology, captures the long-period trend component in the standardized data stream through a temporal convolutional network, and dynamically weights the impact weight of key events by combining an attention mechanism. After fusing the spatial topological relationship with the long-period trend component, it outputs a probability signal of furnace condition evolution.

[0009] The feedback module receives the furnace condition evolution probability signal and generates virtual blast furnace state data based on the physical field simulation model of gas-solid two-phase flow and heat transfer coupling. It then performs differential comparison between the virtual blast furnace state data and the actual collected data to generate parameter calibration signals.

[0010] The early warning module generates graded early warning instructions based on the degree to which the parameter calibration signal crosses a preset threshold.

[0011] In one embodiment of the present invention, the data acquisition module includes a vibration sensor array. The vibration sensor array is arranged with non-uniform density according to the stress gradient distribution of the blast furnace body. The sensor spacing in the throat area is smaller than that in the waist area. Each sensor is configured with a dual-channel acquisition mode to capture the vibration components in the vertical and horizontal directions respectively. After the vibration waveform and the material surface morphology distribution obtained by the ultrasonic scanning device are aligned with the timestamp, multi-source data fusion is performed using a frequency domain convolution kernel to generate an enhanced original feature signal containing the characteristic frequency of the material column collapse. The scanning cycle of the infrared thermal imaging device is dynamically adjusted according to the blast furnace smelting stage. A continuous scanning mode is used in the tapping area to generate a time series map of thermal radiation distribution.

[0012] In one embodiment of the present invention, when the preprocessing module executes the Gaussian process regression algorithm, it dynamically adjusts the noise filtering threshold according to the working state of the blast furnace. When the top pressure fluctuation is detected to exceed the set range, the adaptive covariance function is automatically activated. The kernel parameters of the covariance function are updated online based on the historical noise distribution characteristics. After the denoising process is completed, the standardized data stream is decomposed into multiple scales to extract the low-frequency component containing the material surface descent rate and the high-frequency component reflecting the change of gas flow rate. The standardized data stream with working condition adaptive characteristics is generated by time-frequency domain reconstruction.

[0013] In one embodiment of the present invention, the graph convolutional network in the prediction module constructs a spatial topology graph based on the three-dimensional structure of the blast furnace. The nodes include parameters such as temperature, pressure, material level, and gas composition. The edge weights are dynamically calculated based on the correlation strength between the gas flow rate and the porosity of the material holes. The temporal convolutional network adopts an dilated causal convolutional structure, and its receptive field covers a time span of at least three smelting cycles. When it is detected that the weight allocation of the attention mechanism to key events exceeds a preset ratio, the edge weight reconstruction mechanism of the spatial topology graph is triggered, and the reconstructed topological relationship and trend component are tensor-concatenated and input into the gating fusion unit.

[0014] In one embodiment of the present invention, the physical field simulation model in the feedback module includes a coupled solver for the gas-solid two-phase flow dynamics equation and the radiation heat transfer equation. After receiving the furnace condition evolution probability signal, it starts multi-threaded parallel calculation. The gas phase motion is simulated by the discrete element method to simulate the gas flow field distribution, and the solid phase material motion is tracked by the improved DEM-CFD coupling algorithm to track the ore falling trajectory. The differential comparison process includes establishing a multi-dimensional feature space mapping relationship between actual data and simulation data, generating a parameter calibration signal by calculating the Mahalanobis distance, and encoding the calibration signal into a gradient update vector and injecting it into the network weight adjustment process of the prediction module.

[0015] In one embodiment of the present invention, the early warning module includes a multi-level threshold dynamic generation mechanism. When the parameter calibration signal first touches the primary threshold, the early warning prediction mode is activated. The risk evolution path is estimated by searching similar signal patterns in the historical operating condition database. When the signal continuously crosses the secondary threshold, the emergency response protocol is activated. The protocol includes automatically retrieving equipment maintenance records and spare parts inventory information to generate a maintenance plan. The deviation heat map between the current furnace condition parameters and the standard operating condition is overlaid and displayed through a three-dimensional visualization interface.

[0016] In one embodiment of the present invention, a self-learning optimization module is also included. The self-learning optimization module receives false alarm records and operation feedback signals generated by the early warning module, constructs a loss function by comparing the deviation between the actual furnace condition evolution path and the prediction result, uses a meta-learning framework to jointly optimize the hyperparameters of the graph convolutional network and the temporal convolutional network in the prediction module, and synchronizes the optimized network structure parameters to the simulation calculation engine of the digital twin model to form a closed-loop knowledge evolution link.

[0017] In one embodiment of the present invention, a data quality assessment unit is provided between the data acquisition module and the preprocessing module. The assessment unit monitors the integrity and consistency of the original feature signal in real time. When a sensor failure or data transmission interruption is detected, a missing data compensation mechanism based on generative adversarial network is activated. The compensation mechanism uses the joint distribution characteristics of multimodal data under normal operating conditions to generate a substitute signal, and automatically performs residual correction between the compensation signal and the real signal after the data is recovered.

[0018] In one embodiment of the present invention, the prediction module is configured with a model credibility evaluation mechanism. When the standard deviation of the furnace condition evolution probability signal exceeds a set range, a multi-model voting mechanism is automatically triggered. The voting mechanism simultaneously runs three parallel prediction models, including a spatiotemporal convolutional network, a deep belief network, and a gated recursive unit. The optimal prediction value is selected by comparing the KL divergence of the output results of each model, and the divergence samples are stored in the feature library for subsequent model training.

[0019] This invention also provides an intelligent blast furnace condition trend early warning method, comprising:

[0020] S1: Collect real-time parameters of the blast furnace and generate spatiotemporally correlated original feature signals;

[0021] S2: Receive the original feature signal and filter out noise interference through Gaussian process regression algorithm to form a denoised standardized data stream;

[0022] S3: Receives standardized data stream, extracts spatial topological relationships between blast furnace parameters through graph convolutional network, generates correlation features between gas flow distribution and material surface morphology, captures long-term trend components in standardized data stream through temporal convolutional network, and dynamically weights the impact weights of key events by combining attention mechanism, and outputs a probability signal of furnace condition evolution after fusing spatial topological relationships and long-term trend components.

[0023] S4: Receive the furnace condition evolution probability signal and generate virtual blast furnace state data based on the physical field simulation model of gas-solid two-phase flow and heat transfer coupling. Then, perform differential comparison between the virtual blast furnace state data and the actual collected data to generate parameter calibration signals.

[0024] S5: Generate graded early warning instructions based on the degree to which the parameter calibration signal crosses the preset threshold.

[0025] This invention provides an intelligent blast furnace condition trend early warning system and method. It collects multi-dimensional data through vibration, ultrasonic, and infrared sensors. After noise filtering and standardization, it extracts spatial correlation features between parameters using a graph convolutional network, captures long-term trends using a temporal convolutional network, and dynamically weights key events through an attention mechanism. A digital twin module compares virtual simulation data with actual operating conditions in real time, generating parameter calibration signals to optimize the prediction model. The early warning module outputs operational suggestions in stages based on the degree of deviation between the calibration signal and the threshold, forming a closed loop of "perception-prediction-calibration-decision," ultimately solving the problem of poor adaptability of traditional methods to complex operating conditions. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a system architecture diagram of an intelligent blast furnace condition trend early warning system;

[0028] Figure 2 This is a flowchart of a method for early warning of intelligent blast furnace condition trends. Detailed Implementation

[0029] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0030] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0031] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0032] Please see Figure 1-2The image shows an intelligent blast furnace condition trend early warning system and method according to the present invention. The intelligent blast furnace condition trend early warning system of the present invention includes a data acquisition module, a preprocessing module, a prediction module, a feedback module, and an early warning module. The data acquisition module acquires real-time parameters of the blast furnace and generates spatiotemporally correlated raw feature signals; the preprocessing module receives the raw feature signals and filters out noise interference using a Gaussian process regression algorithm to form a denoised standardized data stream; the prediction module receives the standardized data stream, extracts the spatial topological relationship between blast furnace parameters through a graph convolutional network, generates the correlation features between gas flow distribution and material surface morphology, then captures the long-period trend component in the standardized data stream through a temporal convolutional network, and dynamically weights the influence weight of key events using an attention mechanism, merging the spatial topological relationship and the long-period trend component to output a furnace condition evolution probability signal; the feedback module receives the furnace condition evolution probability signal and generates virtual blast furnace state data based on a physical field simulation model of gas-solid two-phase flow and heat transfer coupling, and performs differential comparison between the virtual blast furnace state data and the actual acquired data to generate a parameter calibration signal; the early warning module generates graded early warning instructions based on the degree to which the parameter calibration signal crosses a preset threshold.

[0033] like Figure 1As shown, the system comprises a complete technology chain consisting of a data acquisition module, a preprocessing module, a prediction module, a feedback module, and an early warning module. The data acquisition module, acting as the system's sensing layer, captures multi-dimensional physical signals during blast furnace operation in real time through vibration sensor arrays, ultrasonic scanning devices, and infrared thermal imaging equipment deployed at different locations within the furnace body. The vibration sensor array employs a non-uniform distribution strategy, focusing on monitoring stress changes in the furnace throat region. Its sampling frequency is dynamically adjusted according to the smelting stage to ensure the capture of characteristic vibration waveforms before the collapse of the burden column. The ultrasonic scanning device operates in pulse-reflection mode, constructing a three-dimensional distribution map of the burden surface morphology by emitting high-frequency sound waves and receiving echo signals, while simultaneously recording the temporal changes in the burden descent rate. The infrared thermal imaging device uses a rotating scanning mechanism to periodically measure the thermal radiation distribution on the furnace surface, particularly employing a continuous scanning mode in the taphole region to generate spatially continuous temperature field data. The raw signals acquired by these heterogeneous sensors are aligned through a timestamp synchronization mechanism, forming spatiotemporally correlated raw characteristic signals, providing a multimodal data foundation for subsequent analysis. The preprocessing module undertakes the core functions of data cleaning and feature enhancement. After receiving the raw feature signals from the data acquisition module, it first performs signal integrity verification to detect sensor failures or abnormal data transmission. For normal data streams, a Gaussian process regression algorithm is used for noise filtering. The covariance function parameters of this algorithm are dynamically optimized according to the current operating state of the blast furnace. For example, when the furnace top pressure fluctuates drastically, it automatically switches to the more robust Matern kernel function and adjusts the length scale parameter based on the historical noise statistical characteristics within the sliding window. The denoised data stream undergoes multi-scale decomposition processing, using wavelet transform to decompose the signal into low-frequency components reflecting the macroscopic movement trend of the material surface and high-frequency components characterizing the transient fluctuations of the gas flow velocity. After each sub-band signal is normalized, a standardized data stream is generated through time-frequency domain reconstruction technology. This data stream not only retains the physical meaning of the original signal but also enhances the expressive dimension of the operating condition characteristics.

[0034] Furthermore, the prediction module, serving as the intelligent analysis hub of the system, focuses on constructing a hybrid spatiotemporal prediction model to accurately predict the evolution trend of furnace conditions. After receiving the standardized data stream output from the preprocessing module, this module simultaneously initiates two computational paths: spatial feature extraction and temporal trend analysis. In the spatial dimension, the graph convolutional network constructs a parameter-related topology graph based on the blast furnace's three-dimensional structure, using parameters such as temperature, pressure, burden height, and gas composition as graph nodes. The edge weights between nodes are calculated in real-time based on the dynamic correlation between gas flow velocity and burden porosity. Through multi-layer graph convolution operations, the network learns the deep correlation features between gas flow distribution and burden morphology. In the temporal dimension, the temporal convolutional network employs a dilated causal convolution structure, with its dilation factor increasing exponentially to form a receptive field covering multiple smelting cycles, effectively capturing the gradual changes in long-term trends. Simultaneously, an attention mechanism dynamically assigns weights to key events in the input sequence (such as charging cycles and tapping operations), strengthening their impact on the prediction results. Spatial topological features and temporal trend components are subjected to tensor splicing and nonlinear transformation through a gated fusion unit, ultimately outputting a probabilistic signal characterizing the future direction of furnace condition evolution. This signal not only includes an assessment of the probability of anomalies occurring but also indicates the time window for risk evolution. The specific implementation of the data acquisition module has undergone technological refinement and innovative design. The vibration sensor array is deployed using a stress gradient-guided non-uniform density configuration, with 6-8 sensor nodes per square meter in the furnace throat region and 3-4 per square meter in the furnace waist region. This differentiated layout effectively matches the stress distribution characteristics generated by material movement within the furnace. Each vibration sensor node integrates a dual-channel acquisition circuit, configured with piezoelectric sensing elements in both vertical and horizontal directions, and synchronously acquires three-dimensional vibration components through orthogonally arranged accelerometers. The acquired raw vibration waveforms are processed by anti-aliasing filtering and then spatiotemporally aligned with the material surface morphology data acquired by the ultrasonic scanning device, specifically through a GPS synchronization clock module to achieve microsecond-level timestamp synchronization. The data fusion stage employs frequency-domain convolution kernel technology, performing convolution operations in the frequency domain on the spectral characteristics of the vibration signal and the ultrasonic material surface data to extract specific frequency band energy features characterizing the risk of material column collapse. The infrared thermal imaging device uses dual-band imaging technology, simultaneously acquiring mid-wave and long-wave infrared spectral data. A multi-spectral fusion algorithm eliminates interference from the oxide layer on the furnace surface on thermal radiation measurements, generating a high-precision temperature distribution map. In the taphole area, the system activates a high-speed scanning mode, increasing the sampling frequency to five times that of the conventional mode, fully recording the transient temperature field changes caused by the molten iron flow. This data provides crucial information for determining the hearth erosion status.

[0035] Specifically, the preprocessing module employs a dynamic optimization mechanism and feature enhancement method. This module integrates a condition sensing unit to monitor the blast furnace's smelting stage and operating status in real time. When the furnace top pressure fluctuation exceeds a set threshold, it automatically triggers a noise filtering parameter adjustment mechanism. In this state, the covariance function of the Gaussian process regression algorithm switches to a composite kernel structure, linearly combined with a radial basis function kernel and a periodic kernel. This structure can capture the transient characteristics of sudden noise while filtering out periodic interference. The hyperparameters of the kernel function are continuously optimized through an online learning mechanism. The marginal likelihood function is calculated using noise samples within a sliding window, and the length scale and signal variance parameters are iteratively updated using the gradient ascent method. The denoised data stream enters a multi-scale analysis stage. Empirical mode decomposition (IMF) adaptively decomposes the signal into multiple intrinsic mode functions (IMFs). Low-frequency IMF components reflecting the macroscopic trend of the material surface movement and high-frequency IMF components characterizing the gas flow velocity pulsation are selected for processing separately. The low-frequency components are smoothed to eliminate random fluctuations, while envelope analysis is performed on the high-frequency components to extract their modulation characteristics. The reorganization process incorporates time-frequency matrix fusion technology, which weights and superimposes the time-frequency spectra of each sub-band signal to generate an enhanced standardized data stream with both time and frequency resolution. This data stream particularly enhances the following features: the first derivative of the material surface descent rate, the second moment statistics of the gas flow velocity change, and the spatial gradient features of the temperature field distribution, providing a high-information-density input feature set for subsequent prediction models.

[0036] In one embodiment of the present invention, the physical field simulation model and the parameter calibration mechanism operate in a coordinated manner. The feedback module constructs a high-precision digital twin model to convert the furnace condition evolution probability signal output by the prediction module into real-time status data of the virtual blast furnace. The physical field simulation model adopts multi-physics coupling solution technology, integrating the gas-solid two-phase flow dynamics equations and radiation heat transfer equations. The gas phase motion simulation is based on the discrete element method, decomposing the gas flow into a discrete particle swarm, and constructing a three-dimensional flow field distribution by calculating the collisions and momentum transfer between particles. The solid phase material motion adopts an improved DEM-CFD coupling algorithm, introducing a viscoplastic contact model into the traditional discrete element framework to accurately simulate the deformation characteristics caused by high-temperature softening during the ore fall. When the simulation calculation starts, the module automatically allocates computing resources, starts independent threads for parallel solution of gas phase and solid phase motion, and realizes real-time data interaction between the two phases through shared memory technology. After the virtual blast furnace status data is generated, the system establishes a multi-dimensional feature space mapping relationship between the actual collected data and the simulation data. Temperature gradient distribution, peak gas flow velocity, and material settling rate are selected as key comparison dimensions. The degree of deviation is assessed by calculating the Mahalanobis distance between the two sets of data in the feature space. This distance value, after normalization, is compared with a preset dynamic tolerance threshold to generate a parameter calibration signal. During the calibration signal encoding process, the system uses the gradient projection method to decompose the multi-dimensional deviation vector into the weight update directions of the graph convolutional network and the temporal convolutional network in the prediction module. A momentum optimization algorithm with a temporal decay factor is injected into the network parameter adjustment process to ensure that the model can quickly respond to the current deviation during calibration while avoiding overfitting caused by short-term fluctuations.

[0037] like Figure 1As shown, the hierarchical decision-making mechanism and visualization assistance function of the early warning module have undergone in-depth technical expansion. The early warning module has a built-in multi-level threshold dynamic generation engine, which automatically adjusts the threshold curve according to the blast furnace smelting cycle, changes in raw material ratio, and equipment operating status. When the parameter calibration signal first touches the primary threshold, the system activates the early warning prediction mode. It retrieves historical records of similar operating conditions within the past year through the graph database, extracts multi-dimensional features including anomaly type, evolution speed, and handling effect, aligns the time series using a dynamic time warping algorithm, and constructs a risk path prediction model based on random forest, outputting the probability distribution of furnace condition deterioration in the next three hours. If the signal continuously crosses the secondary threshold, the module activates the emergency response protocol. First, it calls the maintenance logs and sensor health status data in the equipment management system, and combines them with spare parts inventory information to generate a maintenance plan list with priorities. At the same time, it activates the three-dimensional visualization engine to spatially map and compare the current furnace condition parameters with the standard operating condition benchmark values. It overlays and displays a temperature deviation heat map, stress anomaly areas, and gas flow rate exceeding the standard on the surface of the digital twin model, and distinguishes the risk level through color gradient and flashing frequency. The visual interface also integrates an operation suggestion simulation function. When an operator selects a certain action, the system simulates the furnace condition evolution trend in real time for the next hour after the action is implemented, displaying the predicted changes in key parameters in the form of dynamic curves to help decision-makers evaluate the expected effects of different options. In addition, the module has a built-in feedback learning mechanism that compares and analyzes each early warning response result with the actual furnace condition changes, automatically correcting the weight coefficients in the threshold generation algorithm to form a continuously optimized decision support system.

[0038] Furthermore, the self-learning optimization module constructs a closed-loop learning link from early warning feedback to model evolution. This module continuously receives false alarm cases recorded by the early warning module, valid alarms confirmed by operators, and data on subsequent handling effects. It constructs multi-dimensional feature vectors to characterize the environmental context of each early warning event, including auxiliary parameters such as the fluctuation range of raw material composition, the percentile value of equipment operating time, and environmental temperature and humidity conditions. To address the prediction bias problem, the module designs a composite loss function, decomposing the prediction error of the furnace condition evolution probability signal into three components: spatial feature error, temporal trend error, and correlation weight error. It calculates the cosine similarity loss between the gas flow-material surface correlation features output by the graph convolutional network and the actual observed values, the mean square error loss between the predicted trend of the temporal convolutional network and the true value, and the KL divergence loss between the attention mechanism weight distribution and expert-annotated key events. Based on this loss function, a meta-learning framework is used to perform two-layer optimization of the prediction model: in the inner loop, the network weights are updated using the current batch of data; in the outer loop, the network hyperparameters are optimized through cross-task knowledge transfer, including the neighborhood aggregation radius of the graph convolutional layer, the expansion coefficient sequence of the temporal convolutional network, and the temperature coefficient of the attention mechanism. The optimized network parameters are synchronized to the simulation calculation engine of the digital twin model. Specifically, through the parameter mapping interface, the spatial association rules learned by the graph convolutional network are transformed into the coupling coefficients of the gas flow and material surface movement in the physical field simulation model, and the long-period trend features captured by the temporal convolutional network are embedded into the simulation initial condition generation algorithm, realizing the knowledge co-evolution of the prediction model and the simulation model. This process forms a positive feedback loop, enabling the system to continuously improve its adaptability to complex working conditions during continuous operation.

[0039] like Figure 2 The diagram illustrates an intelligent blast furnace condition trend early warning method according to the present invention. S1: Collect real-time parameters of the blast furnace and generate spatiotemporally correlated original feature signals; S2: Receive the original feature signals and filter out noise interference using a Gaussian process regression algorithm to form a denoised standardized data stream; S3: Receive the standardized data stream, extract the spatial topological relationship between blast furnace parameters using a graph convolutional network, generate the correlation features between gas flow distribution and material surface morphology, then capture the long-period trend component in the standardized data stream using a temporal convolutional network, and dynamically weight the influence weight of key events using an attention mechanism, merging the spatial topological relationship with the long-period trend component to output a furnace condition evolution probability signal; S4: Receive the furnace condition evolution probability signal and generate virtual blast furnace state data based on a physical field simulation model of gas-solid two-phase flow and heat transfer coupling, and perform differential comparison between the virtual blast furnace state data and the actual collected data to generate a parameter calibration signal; S5: Generate graded early warning instructions based on the degree to which the parameter calibration signal crosses a preset threshold.

[0040] Furthermore, the data quality assurance system adds an intelligent data repair unit between the data acquisition module and the preprocessing module. This unit monitors the integrity indicators of the original feature signals in real time, including sensor sampling rate stability, data packet temporal continuity, and cross-modal data alignment accuracy. When a specific sensor node failure is detected or the data transmission delay exceeds the tolerance threshold, a data compensation mechanism based on a deep generative adversarial network is immediately activated. The generator network adopts a U-Net architecture, inputting the joint distribution characteristics of multi-sensor data under normal operating conditions. Through a spatial attention gating mechanism, it learns the implicit correlation between vibration waveforms, material surface morphology, and thermal radiation distribution, and outputs a synthetic signal that conforms to the current operating context. The discriminator network constructs a multi-scale feature pyramid to distinguish between real and generated signals from three dimensions: time-domain waveform, frequency-domain energy distribution, and spatiotemporal correlation. During adversarial training, a course learning strategy is introduced to gradually increase the proportion of missing data to improve the robustness of the generator. Once data transmission is restored, the system automatically performs residual correction between the compensated signal and the real signal: first, it calculates the mean square error of both signals within a sliding time window; if the error exceeds the allowable range, it triggers a Kalman filter algorithm to dynamically correct the subsequently generated compensated signal; simultaneously, compensated data from abnormal periods are marked as low-confidence samples for weighted processing during subsequent model retraining. Furthermore, the unit incorporates a built-in sensor health assessment model, which analyzes the signal-to-noise ratio, zero-drift, and spectral characteristic trends of signals at each node to predict sensor performance degradation risks in advance and generate maintenance warnings for the equipment management system.

[0041] This invention discloses an intelligent blast furnace condition trend early warning system and method. It collects multi-dimensional data through vibration, ultrasonic, and infrared sensors. After noise filtering and standardization, it extracts spatial correlation features between parameters using a graph convolutional network, captures long-term trends using a temporal convolutional network, and dynamically weights key events through an attention mechanism. A digital twin module compares virtual simulation data with actual operating conditions in real time, generating parameter calibration signals to optimize the prediction model. The early warning module outputs operational suggestions in stages based on the degree of deviation between the calibration signal and the threshold, forming a closed loop of "perception-prediction-calibration-decision," ultimately solving the problem of poor adaptability of traditional methods to complex operating conditions.

[0042] Therefore, the intelligent blast furnace condition trend early warning system and method of the present invention solves the problem of delayed early warning caused by the reliance on a single parameter, static model and human experience in traditional blast furnace condition early warning methods.

[0043] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. An intelligent blast furnace condition trend early warning system, characterized in that, include: The data acquisition module collects real-time parameters of the blast furnace and generates spatiotemporally correlated raw feature signals. The data acquisition module captures multi-dimensional physical signals during blast furnace operation in real time through vibration sensor arrays, ultrasonic scanning devices, and infrared thermal imaging equipment deployed at different locations in the furnace body. The raw signals collected by these sensors are aligned through a timestamp synchronization mechanism to form the spatiotemporally correlated raw feature signals. The preprocessing module receives the original feature signal and filters out noise interference through a Gaussian process regression algorithm to form a denoised standardized data stream. The prediction module receives the standardized data stream, extracts the spatial topological relationship between blast furnace parameters through a graph convolutional network, generates the correlation features between gas flow distribution and material surface morphology, captures the long-period trend component in the standardized data stream through a temporal convolutional network, and dynamically weights the influence weight of key events by combining an attention mechanism. After fusing the spatial topological relationship with the long-period trend component, the module outputs a probability signal of furnace condition evolution. The feedback module receives the furnace condition evolution probability signal and generates virtual blast furnace state data based on the physical field simulation model of gas-solid two-phase flow and heat transfer coupling. The feedback module then performs differential comparison between the virtual blast furnace state data and the actual collected data to generate a parameter calibration signal. The early warning module generates graded early warning instructions based on the degree to which the parameter calibration signal crosses a preset threshold.

2. The intelligent blast furnace condition trend early warning system according to claim 1, characterized in that, The data acquisition module includes a vibration sensor array, which is arranged with a non-uniform density according to the stress gradient distribution of the blast furnace body. The sensor spacing in the throat area is smaller than that in the waist area. Each sensor is configured with a dual-channel acquisition mode to capture the vibration components in the vertical and horizontal directions respectively. After the vibration waveform and the material surface morphology distribution obtained by the ultrasonic scanning device are aligned with the timestamp, multi-source data fusion is performed using a frequency domain convolution kernel to generate an enhanced original feature signal containing the characteristic frequency of the material column collapse. The scanning cycle of the infrared thermal imaging device is dynamically adjusted according to the blast furnace smelting stage. A continuous scanning mode is used in the tapping area to generate a time series map of thermal radiation distribution.

3. The intelligent blast furnace condition trend early warning system according to claim 1, characterized in that, When executing the Gaussian process regression algorithm, the preprocessing module dynamically adjusts the noise filtering threshold according to the blast furnace operating status. When the top pressure fluctuation exceeds the set range, the adaptive covariance function is automatically activated. The kernel parameters of the covariance function are updated online based on historical noise distribution characteristics. After completing the denoising process, the standardized data stream is decomposed into multiple scales to extract the low-frequency component containing the material surface descent rate and the high-frequency component reflecting the gas flow rate change. The standardized data stream with operating condition adaptive characteristics is generated through time-frequency domain reconstruction.

4. The intelligent blast furnace condition trend early warning system according to claim 1, characterized in that, The graph convolutional network in the prediction module constructs a spatial topology graph based on the three-dimensional structure of the blast furnace. The nodes include parameters such as temperature, pressure, material level, and gas composition. The edge weights are dynamically calculated based on the correlation strength between the gas flow rate and the porosity of the material holes. The temporal convolutional network adopts a dilated causal convolutional structure, and its receptive field covers a time span of at least three smelting cycles. When the attention mechanism is detected to assign more weights to key events than a preset ratio, the edge weight reconstruction mechanism of the spatial topology graph is triggered. The reconstructed topological relationships and trend components are then tensor-concatenated and input into the gating fusion unit.

5. The intelligent blast furnace condition trend early warning system according to claim 1, characterized in that, The physical field simulation model in the feedback module includes a coupled solver for the gas-solid two-phase flow dynamics equation and the radiation heat transfer equation. After receiving the furnace condition evolution probability signal, it starts multi-threaded parallel calculation. The gas phase motion uses the discrete element method to simulate the gas flow field distribution, and the solid phase material motion tracks the ore falling trajectory through an improved DEM-CFD coupling algorithm. The differential comparison process includes establishing a multi-dimensional feature space mapping relationship between actual data and simulation data, generating parameter calibration signals by calculating Mahalanobis distance, and encoding the calibration signals as gradient update vectors into the network weight adjustment process of the prediction module.

6. The intelligent blast furnace condition trend early warning system according to claim 1, characterized in that, The early warning module includes a multi-level threshold dynamic generation mechanism. When the parameter calibration signal first touches the primary threshold, the early warning prediction mode is activated. The risk evolution path is estimated by searching the historical operating condition database for similar signal patterns. When the signal continuously crosses the secondary threshold, the emergency response protocol is activated. The protocol includes automatically retrieving equipment maintenance records and spare parts inventory information to generate a maintenance plan, and displaying the deviation thermal map between the current furnace condition parameters and the standard operating condition through a three-dimensional visualization interface.

7. The intelligent blast furnace condition trend early warning system according to claim 1, characterized in that, It also includes a self-learning optimization module, which receives false alarm records and operation feedback signals generated by the early warning module, constructs a loss function by comparing the deviation between the actual furnace condition evolution path and the prediction result, uses a meta-learning framework to jointly optimize the hyperparameters of the graph convolutional network and the temporal convolutional network in the prediction module, and synchronizes the optimized network structure parameters to the simulation calculation engine of the digital twin model to form a closed-loop knowledge evolution link.

8. The intelligent blast furnace condition trend early warning system according to claim 1, characterized in that, A data quality assessment unit is provided between the data acquisition module and the preprocessing module. The assessment unit monitors the integrity and consistency of the original feature signals in real time. When sensor failure or data transmission interruption is detected, a missing data compensation mechanism based on generative adversarial networks is activated. The compensation mechanism uses the joint distribution characteristics of multimodal data under normal operating conditions to generate a substitute signal, and automatically performs residual correction between the compensation signal and the real signal after data recovery.

9. The intelligent blast furnace condition trend early warning system according to claim 1, characterized in that, The prediction module is configured with a model credibility evaluation mechanism. When the standard deviation of the furnace condition evolution probability signal exceeds the set range, a multi-model voting mechanism is automatically triggered. The voting mechanism simultaneously runs three parallel prediction models, including a spatiotemporal convolutional network, a deep belief network, and a gated recursive unit. The optimal prediction value is selected by comparing the KL divergence of the output results of each model, and the divergence samples are stored in the feature library for subsequent model training.

10. A method for intelligent early warning of blast furnace condition trends, characterized in that, The intelligent blast furnace condition trend early warning system according to any one of claims 1-9 includes: S1: Collect real-time parameters of the blast furnace and generate spatiotemporally correlated original feature signals; S2: Receive the original feature signal and filter out noise interference through Gaussian process regression algorithm to form a denoised standardized data stream; S3: Receive the standardized data stream, extract the spatial topological relationship between blast furnace parameters through graph convolutional network, generate the correlation features between gas flow distribution and material surface morphology, capture the long-period trend component in the standardized data stream through temporal convolutional network, and dynamically weight the influence weight of key events by combining the attention mechanism, and output the furnace condition evolution probability signal after fusing the spatial topological relationship with the long-period trend component. S4: Receive the furnace condition evolution probability signal and generate virtual blast furnace state data based on the physical field simulation model of gas-solid two-phase flow and heat transfer coupling, and perform differential comparison between the virtual blast furnace state data and the actual collected data to generate parameter calibration signal. S5: Generate a graded early warning command based on the degree to which the calibration signal crosses a preset threshold according to the parameters.

Images