Integrated monitoring system for synergic control of heating furnace protective atmosphere and dew point

Abstract

This invention discloses an integrated monitoring system for the coordinated control of protective atmosphere and dew point in a heating furnace, belonging to the field of metal processing and intelligent manufacturing technology. It includes a data acquisition module, an alignment processing module, an analytical extraction module, a flow field reconstruction module, an identification and positioning module, a state fusion module, a virtual reconstruction module, an evaluation and judgment module, an exploration and decision-making module, an instruction generation module, and a feedback optimization module. This invention can dynamically reflect the velocity distribution and turbulence changes, thus overcoming the limitations of traditional single-point measurements in obtaining spatial distribution information. It improves the accuracy and response speed of fault detection, enhances the safety of system operation, effectively ensures the continuity and reliability of key control variables, avoids control interruptions due to sensor failure, and achieves a coordinated control effect with no overshoot, fast response, and strong robustness.

Description

Technical Field

[0001] This invention relates to the fields of metal machining and intelligent manufacturing technology, and in particular to an integrated monitoring system for the coordinated control of protective atmosphere and dew point in a heating furnace. Background Technology

[0002] In metallurgy, heat treatment, and materials processing, heating furnaces are core process equipment. The stability of their internal protective atmosphere and the accuracy of dew point control directly determine the surface quality, microstructure, and product consistency of workpieces. However, in actual operation, the flow of atmosphere within the furnace often exhibits strong non-uniformity and dynamic fluctuations. Local leaks, atmosphere disturbances, and changes in heat load can easily lead to imbalances in atmosphere concentration distribution, resulting in oxidation, decarburization, or surface defects. Dew point, as a key parameter characterizing moisture content, is significantly affected by sensor drift, response hysteresis, and interference from high-temperature environments, posing a risk of data distortion or even failure. Existing technologies mostly rely on single-point sensors or empirical models to independently control the atmosphere and dew point, lacking real-time sensing capabilities of the three-dimensional flow state inside the furnace, making it difficult to accurately locate abnormal atmosphere distribution. Once the dew point sensor malfunctions, the system often cannot provide reliable alternative data, leading to control strategy failure. Furthermore, traditional control methods often employ fixed parameters or simple feedback mechanisms, which are ill-suited to address the strong coupling, multivariable, and nonlinear dynamic characteristics of heating furnace operation, making the contradiction between control response speed and stability particularly prominent.

[0003] To overcome the aforementioned shortcomings, an integrated monitoring system for the coordinated control of protective atmosphere and dew point in heating furnaces has been proposed in recent years. This system introduces a multi-physics coupled sensing architecture, deploying distributed velocity, temperature, and dew point sensors in key areas of the furnace. Combined with computational fluid dynamics models, it achieves dynamic reconstruction of velocity distribution and turbulence changes within the furnace. Compared to traditional single-point detection methods, this system can effectively acquire spatial distribution information of the atmosphere state, improving the accuracy and response speed of fault detection. Simultaneously, the system incorporates a redundant data fusion and self-diagnosis module, which can generate reliable alternative values ​​based on historical data and spatial correlation models when some sensors malfunction or fail, thereby ensuring the continuity and reliability of key control variables and preventing control interruptions due to sensor failures. Summary of the Invention

[0004] The purpose of this invention is to provide an integrated monitoring system for the coordinated control of protective atmosphere and dew point in a heating furnace, in order to solve the problems of existing technologies that cannot dynamically reflect the flow velocity distribution and turbulence changes in the furnace, lack spatial distribution information, have low fault detection accuracy and response speed, are prone to control interruption due to sensor failure, and are difficult to achieve coordinated control effects with no overshoot, fast response and strong robustness.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An integrated monitoring system for the coordinated control of protective atmosphere and dew point in a heating furnace includes a data acquisition module, an alignment processing module, an analytical extraction module, a flow field reconstruction module, an identification and positioning module, a state fusion module, a virtual reconstruction module, an evaluation and judgment module, an exploration and decision-making module, an instruction generation module, and a feedback optimization module. The data acquisition module collects real-time sensor data on the furnace atmosphere propagation characteristics, thermal state, and composition information from various sources. The alignment processing module aligns the time of the multi-source sensor data and constructs a standardized observation vector. The analytical extraction module demodulates and performs frequency domain analysis on the swept-frequency response signal from the multi-source sensor data and extracts the sound wave propagation characteristics. The flow field reconstruction module reconstructs the three-dimensional flow velocity distribution of the protective atmosphere inside the furnace based on the sound wave propagation characteristics. The flow field layout and turbulence intensity distribution are analyzed. The identification and localization module performs time-series evolution analysis on the reconstructed flow field, identifies various abnormal modes, and locates potential leakage points or disturbance sources. The state fusion module constructs a unified state vector and introduces historical control behavior sequences to form a high-dimensional state vector. The virtual reconstruction module generates dew point data in real time based on various sensor data. The evaluation and judgment module dynamically evaluates the fused high-dimensional state vector and identifies various critical states. The exploration and decision-making module learns and optimizes the control strategy using the current state and historical feedback. The instruction generation module generates multi-variable collaborative control instructions according to the control strategy. The feedback optimization module collects the system response data after execution in real time and compares it with the prediction results to dynamically correct the control strategy.

[0006] This invention constructs a complete technology chain from data perception, feature extraction, flow field reconstruction, anomaly localization to intelligent decision-making and closed-loop optimization through the collaborative work of a data acquisition module, alignment processing module, analytical extraction module, flow field reconstruction module, identification and localization module, state fusion module, virtual reconstruction module, evaluation and judgment module, exploration and decision-making module, instruction generation module, and feedback optimization module. Compared with existing technologies, this system can not only acquire the three-dimensional velocity distribution and turbulence intensity of the furnace atmosphere in real time, breaking through the spatial limitations of single-point measurement, but also generate reliable alternative data through the virtual reconstruction module when the dew point sensor malfunctions, avoiding control interruption. At the same time, the reinforcement learning-based exploration and decision-making module and feedback optimization module realize continuous adaptive improvement of the control strategy, improving the system's response speed and control stability under strongly nonlinear and multivariable coupled conditions, ultimately achieving a collaborative control effect with no overshoot and strong robustness.

[0007] As a further aspect of the present invention, the data acquisition module collects various sensor data at key locations of the heating furnace through acoustic arrays, infrared spectral sensors, oxygen potential sensors, temperature field sensors, pressure sensors, and dew point sensors.

[0008] As a further aspect of the present invention, the specific steps of the alignment processing module in aligning multi-source sensor data time and constructing a standardized observation vector are as follows: S1.1: Retain the original sampled value and local timestamp of each channel's sensor data, and bind the original sampled value and local timestamp of each channel's sensor data to form the time series of each channel. Then, collect the original time of each sensor channel at two synchronization moments, and obtain the clock scale correction coefficient and time shift correction amount of the corresponding sensor channel by fitting the synchronization reference point. Then, perform linear correction on the local timestamp of each channel's sensor data. S1.2: After completing the time correction, a unified time axis is established. For any sensing channel, if there is no corresponding sampling point at any time on the unified time axis, linear interpolation is performed between two adjacent corrected sampling points to obtain the alignment value of the sensing channel on the unified time axis. Then, the sliding window of a preset size is moved step by step, and the mean and standard deviation within each sliding window are calculated. S1.3: Based on the mean of each step, the corresponding sensor data of each sensor channel is centered, and then the standard deviation is used to unify the scale to obtain the standardized result. Finally, the standardized values ​​of the sensor data of all sensor channels at the same time are concatenated column by column to form a unified observation vector.

[0009] This solution addresses the data asynchrony issues caused by different sampling frequencies, inconsistent start times, and local clock drift in multi-source sensor data through steps such as timestamp binding, clock scale correction, linear interpolation alignment, sliding window standardization, and observation vector splicing performed by the alignment processing module. This process ensures accurate alignment and dimensional consistency of the input data on a unified time axis for subsequent acoustic feature extraction and flow field reconstruction, eliminating spurious fluctuations and spatial misjudgments caused by time deviations. This lays a reliable data foundation for high-precision three-dimensional velocity distribution reconstruction and anomaly localization, thereby improving the overall accuracy and robustness of the system's detection.

[0010] As a further aspect of the present invention, the detailed steps of the analytical extraction module in demodulating and performing frequency domain analysis on the swept frequency response signal in the multi-source sensor data are as follows: S2.1: The parsing and extraction module transmits a preset sweep frequency encoded signal through the acoustic wave array. Then, the receiver segments and samples the original echoes of each receiving channel to establish a discrete signal sequence and construct a corresponding reference sweep frequency excitation. After that, the original echoes and the reference sweep frequency signal are conjugate mixed to obtain the demodulated intermediate frequency or baseband components. S2.2: After performing sliding smoothing or low-pass filtering on the demodulated intermediate frequency or baseband components, the corresponding baseband response is generated. Based on the smoothed baseband responses, the corresponding analytical signals are constructed. Based on the established analytical signals, the corresponding instantaneous amplitude and instantaneous phase are obtained. Then, the cross-correlation function between each receiving channel and its reference sweep frequency excitation is calculated. S2.3: Based on the cross-correlation function, calculate the correlation value at different times, extract the time shift corresponding to the correlation peak, and calculate the flight time of the acoustic wave in each receiving channel according to the current sampling period of the acoustic array. Then, perform a discrete Fourier transform on the smoothed baseband response, extract the spectral amplitude from the transformed baseband response, and calculate the propagation attenuation at the current frequency according to the known transmission reference spectrum. S2.4: If the attenuation is linearly related to the propagation distance in any frequency band, calculate the corresponding average propagation attenuation in the current frequency band to obtain the comprehensive attenuation coefficient index of the corresponding receiving channel. Then, extract the phase spectrum from the baseband response after transformation of each receiving channel, and perform phase expansion on each extracted phase spectrum to obtain a continuous phase curve. Subsequently, calculate the phase offset relative to the reference spectrum and calculate the average phase offset in each frequency band. Then, integrate the sound wave flight time, propagation attenuation and phase change extracted from each receiving channel into sound wave propagation characteristics.

[0011] The above-mentioned scheme performs a series of processing steps on the swept frequency response signal through a parsing extraction module, including conjugate mixing demodulation, sliding smoothing, cross-correlation analysis, discrete Fourier transform, and phase expansion. This allows for the robust extraction of key propagation features such as sound wave flight time, propagation attenuation, and phase change from the complex acoustic response within the furnace. This method overcomes the interference of high temperature, strong noise, and multipath effects on the acoustic signal. The extracted acoustic propagation features have a high signal-to-noise ratio and high consistency, providing accurate acoustic observation information for the subsequent flow field reconstruction module and improving the reconstruction accuracy of velocity distribution and turbulence intensity.

[0012] As a further aspect of the present invention, the specific steps of the flow field reconstruction module in reconstructing the three-dimensional velocity distribution field and turbulence intensity distribution of the protective atmosphere inside the furnace are as follows: S3.1: Based on the geometric dimensions of the furnace, the inner wall boundary, and the layout of the acoustic array, construct the three-dimensional computational domain to be reconstructed. Then, discretize the internal space of the furnace into multiple sets of non-overlapping voxel units. Then, between the acoustic transceiver arrays arranged on both sides or around the furnace, any pair of transmitters and receivers constitutes an acoustic propagation path. Calculate the reciprocal of the propagation speed of each acoustic propagation path and use it as the slowness of the corresponding voxel. S3.2: Calculate the equivalent length of each propagation path through each voxel unit, and calculate the measured propagation delay of each sound wave propagation path. Then, use the propagation delay corresponding to each sound wave propagation path as the observation vector, and combine the slowness of all voxel units into the parameter vector to be estimated, so as to generate a unified linear inversion model. S3.3: With the goal of minimizing the time delay error of each sound wave propagation path, a regularization method is used to invert the slowness field of each voxel unit. At the same time, a stable three-dimensional slowness distribution is obtained by jointly solving the least squares and smoothing constraints. The velocity value in the corresponding voxel unit is obtained by calculating the reciprocal of the slowness of each voxel unit, and a three-dimensional velocity distribution is constructed by the velocity values ​​of each voxel unit. S3.4: Statistically analyze the fluctuation range of the velocity of each voxel unit at multiple consecutive reconstruction times, and use the relative fluctuation amount as the turbulence intensity of the corresponding voxel unit. Then, bind the spatial coordinates, velocity value and turbulence intensity value of each voxel to form a complete three-dimensional flow field reconstruction result. Through mapping, display the velocity level, flow direction change and turbulence concentration area in the three-dimensional flow field.

[0013] In the above scheme, the present invention constructs a three-dimensional computational domain and discrete voxel units through a flow field reconstruction module, establishes a linear inversion model of sound wave propagation path and slowness, and uses a regularization method combined with least squares and smoothing constraints to solve the problem. This allows for the stable reconstruction of the three-dimensional velocity distribution and turbulence intensity of each voxel unit inside the furnace. This module not only realizes the quantitative conversion from sound wave flight time to velocity field, but also obtains the spatial distribution of turbulence intensity through time-series fluctuation statistics. This enables operators to intuitively identify high and low velocity regions, changes in flow direction, and turbulence concentration areas, providing intuitive and reliable spatial information support for the accurate location of leakage points and disturbance sources.

[0014] As a further aspect of the present invention, the specific steps of the identification and positioning module in identifying various abnormal patterns and locating potential leakage points or disturbance sources are as follows: S4.1: Store the three-dimensional flow field results obtained from reconstruction at different times in chronological order, and extract the local state variables corresponding to the anomaly identification in each spatial voxel unit to form a time-series analysis object. Then calculate the local baseline of the state characteristics of each voxel unit at each time, and calculate the comprehensive deviation of each voxel unit based on the obtained local baseline at each time. S4.2: Detect whether the overall deviation of each voxel unit is greater than the preset deviation threshold multiple times. If it is greater and the number of consecutive exceedances is higher than the preset persistence judgment threshold, it indicates that the corresponding voxel unit is a candidate abnormal voxel. Count the number of candidate abnormal voxels and construct a set of candidate abnormal voxels. S4.3: Establish an adjacency graph for candidate anomalous voxels according to their spatial location, and expand the candidate anomalous voxels to form multiple anomalous clusters. At the same time, calculate the average anomalous intensity of each anomalous cluster, then calculate the center coordinates of each anomalous cluster by weighting, and calculate its spatial expansion radius based on the center coordinates, average anomalous intensity and comprehensive deviation of each anomalous cluster, and obtain the influence range of each anomalous cluster. S4.4: Map the center coordinates, influence range and average anomaly intensity of the anomaly cluster back to the three-dimensional flow field, generate marker information at the corresponding spatial location, and calculate the alarm score of each anomaly cluster. If the alarm score is higher than the preset comprehensive alarm index and the average number of continuous over-limits of the voxel units in the anomaly cluster is higher than the preset continuous upper limit, an early warning is triggered, and the corresponding spatial location is marked in red or highlighted in the three-dimensional flow field, and the early warning information is output.

[0015] This solution uses an identification and positioning module to perform time-series evolution analysis of the three-dimensional flow field, calculate the comprehensive deviation of each voxel unit, and form anomaly clusters through adjacency graph connectivity expansion. Finally, it triggers early warning and spatial positioning through alarm scoring and continuous over-limit judgment. This module can automatically identify minute and continuous atmospheric anomaly patterns, effectively distinguishing random fluctuations from real leaks or disturbance sources, avoiding the problems of false alarms and missed alarms that are easy to occur with traditional threshold methods. At the same time, it marks or highlights the abnormal areas in the three-dimensional flow field, improving the response speed and positioning accuracy of fault detection, and enhancing the safety of the heating furnace operation.

[0016] As a further aspect of the present invention, the specific steps by which the virtual reconstruction module generates dew point data in real time based on various sensor data are as follows: S5.1: When any dew point sensor malfunctions or fails, the corresponding dew point sensor malfunction score is calculated based on the rolling median baseline formed by the historical effective dew point values ​​and the reference dew point value. If the malfunction score is higher than the preset threshold, the dew point data at that moment is determined to be abnormal or failed, and that time period is marked as the interval to be compensated. At the same time, the output of the dew point sensor is stopped, and synchronous data at the same moment is extracted from the other corresponding sensors, including the furnace infrared spectral characteristics, oxygen potential signal, furnace temperature distribution, furnace pressure fluctuation, and historical effective dew point sequence. S5.2: Encode the features of each synchronization data, add a random noise vector, and construct a joint input vector. Then, send the joint input vector to the condition generator. The condition generator outputs the candidate dew point data of the current missing time or missing interval point by point according to the time window. Then, send the candidate dew point data to the discriminator for authenticity evaluation and generate a confidence score for each candidate dew point data. S5.3: If the credibility score is higher than the preset threshold, it indicates that the corresponding candidate dew point data is acceptable. At the same time, the allowable range of dew point data and the maximum allowable adjacent variation are set, and the physical constraint penalty term of each dew point data at the current time is calculated. If the physical constraint penalty term is higher than the preset intervention threshold, the candidate dew point data at the current time is compressed and corrected. After completing the discrimination and physical correction, the candidate dew point data is weighted and fused according to the credibility score and the physical penalty value to generate the final compensated dew point value.

[0017] Here, a virtual reconstruction module utilizes multi-source synchronous data such as historical dew point baseline, infrared spectrum, oxygen potential, furnace temperature, and furnace pressure when the dew point sensor malfunctions. Combined with the adversarial learning mechanism of the condition generator and discriminator, it generates highly reliable compensated dew point data in real time. Furthermore, it corrects and weights the data through physical constraint penalty terms. This module effectively solves the problem of control interruption or data loss caused by the failure of a single dew point sensor in traditional systems. It ensures the continuity and reliability of key dew point variables, enabling the collaborative control system to maintain normal operation under sensor failure conditions, thereby improving the system's fault tolerance and engineering practicality.

[0018] As a further aspect of the present invention, the steps of the exploratory decision module learning and optimizing the control strategy are as follows: S6.1: At each decision moment, collect the current operating state and historical feedback information, organize them into a control state vector, and output the initial control action according to the current policy network based on the current control state vector, and add random exploration noise to perturb the initial control action; S6.2: Send the actions generated by the disturbance to the controller, and calculate the immediate reward based on the actual response after the controller executes the action. Then, combine the current state, action, reward and the next state into a time series sample, calculate the target Q value based on the current time series sample, calculate the time series difference loss, and store the historical interaction samples into the experience pool in order. S6.3: Extract multiple samples from the experience pool and calculate the corresponding batch average loss based on the temporal difference loss of each sample. If the batch average loss is higher than the preset loss value, update the policy network parameters and generate the execution action again. Continue to iterate and update until the batch average loss is lower than the preset loss value. S6.4: The statistical policy network, after training, generates control policies at different times and calculates their average loss and average reward. If the average loss is lower than the preset threshold and the preset stability condition is met, the policy corresponding to the current parameters is output as the optimal control policy.

[0019] The above scheme constructs a state-action-reward closed-loop training mechanism based on reinforcement learning through an exploration decision module. It introduces random exploration noise to avoid the strategy getting trapped in local optima, and uses experience pool replay and batch average loss evaluation to stabilize the training process. After the average loss and stability conditions are met, the optimal control strategy is output. This module enables the control strategy to adapt to the nonlinearity, strong coupling and multivariate dynamic changes in the operation of the heating furnace, overcoming the shortcomings of slow response and easy overshoot of traditional fixed parameter or simple feedback control. It achieves a cooperative control effect with fast response, no overshoot and strong robustness.

[0020] As a further aspect of the present invention, the control state vector in S6.1 specifically includes the target deviation of the current controlled object, the current atmosphere state deviation, the current pressure state deviation, the current temperature state deviation, and a summary of historical control feedback. The present invention provides a comprehensive and compact state representation for the reinforcement learning agent by constructing the control state vector from the target deviation of the current controlled object, the atmosphere state deviation, the pressure state deviation, the temperature state deviation, and the summary of historical control feedback. This state vector design takes into account both immediate deviations and historical control effects, enabling the decision-making module to accurately perceive the cumulative impact of the current process conditions and previous control behaviors, thereby learning a more forward-looking and coordinated control strategy, further improving the accuracy and dynamic response performance of the coordinated control of protective atmosphere and dew point.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves high-precision characterization of the internal atmosphere and dew point state of a heating furnace by constructing a unified processing framework for multi-source sensor data. First, the system binds the original sampled values ​​and timestamps of the multi-source sensor data to construct the time series of each channel. Through clock correction and linear interpolation, sensor data with different sampling frequencies and start times are precisely aligned on a unified time axis. Based on this, a sliding window is used to calculate the mean and standard deviation of the data for each channel, completing data centralization and scale unification to form a standardized observation vector. Simultaneously, based on an acoustic array, the system transmits and receives swept-frequency signals, sequentially performing demodulation, filtering, and frequency domain analysis on the acquired signals to extract key acoustic features such as sound wave flight time, propagation attenuation, and phase changes, constructing a complete set of sound wave propagation features. A three-dimensional computational domain is established in conjunction with the actual furnace structure. The atmospheric slowness field is reconstructed using tomographic inversion, and then converted into a velocity distribution. Turbulence intensity is calculated by combining time-series fluctuation information, ultimately forming a three-dimensional flow field model that dynamically reflects the velocity distribution and turbulence changes.

[0022] 2. This invention further enhances the system's capabilities in anomaly detection and sensor fault tolerance. By performing time-series analysis on the three-dimensional flow field, the system can automatically identify anomalous voxels of flow velocity and turbulence intensity, and accurately locate anomalous regions using a spatial clustering algorithm. This overcomes the limitations of traditional single-point measurement in obtaining spatial distribution information, improving the accuracy and response speed of fault detection. When the dew point sensor malfunctions, the system utilizes multi-source sensor information to construct a conditional generation model, reconstructs missing dew point data, and achieves correction and fusion through a discriminant network and physical constraint mechanism. This effectively ensures the continuity and reliability of key control variables, avoids control interruption due to sensor failure, and enhances the overall safety and robustness of the system operation.

[0023] 3. This invention constructs a collaborative control strategy based on reinforcement learning. It optimizes control decisions through a closed-loop training mechanism consisting of system state, control actions, and reward feedback. After meeting the convergence conditions, it outputs the optimal control strategy, achieving a collaborative control effect of no overshoot, fast response, and strong robustness for protective atmosphere and dew point. This control strategy can adapt to the nonlinearity, strong coupling, and multivariate disturbances in the operation of the heating furnace, overcoming the inherent contradiction between dynamic response and stability in traditional fixed parameter or simple feedback mechanisms. Through the complete technical chain from data fusion, flow field reconstruction, anomaly localization to intelligent decision-making, this invention improves the automation level and process assurance capability of collaborative control of protective atmosphere and dew point in heating furnaces. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0025] Figure 1 This is a system block diagram of an integrated monitoring system for the coordinated control of protective atmosphere and dew point in a heating furnace, as proposed in this invention. Figure 2 This is a flowchart illustrating the acoustic propagation feature extraction process of an integrated monitoring system for coordinated control of protective atmosphere and dew point in a heating furnace, as proposed in this invention. Figure 3 This is a flowchart illustrating the three-dimensional flow field construction of an integrated monitoring system for the coordinated control of protective atmosphere and dew point in a heating furnace, as proposed in this invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0027] Example 1, referring to Figure 1-3An integrated monitoring system for the coordinated control of protective atmosphere and dew point in a heating furnace includes a data acquisition module, an alignment processing module, an analysis and extraction module, a flow field reconstruction module, an identification and positioning module, a state fusion module, a virtual reconstruction module, an evaluation and judgment module, an exploration and decision-making module, an instruction generation module, and a feedback optimization module.

[0028] The data acquisition module collects real-time sensor data on the propagation characteristics, thermal state, and composition information of the atmosphere inside the furnace; the alignment processing module aligns the time of the multi-source sensor data and constructs a standardized observation vector.

[0029] Specifically, the original sampled values ​​and local timestamps of each sensor data channel are retained, and the original sampled values ​​and local timestamps of each sensor data channel are bound together to form the time series of each channel. Then, the original times of each sensor channel at two synchronization moments are collected, and the clock scale correction coefficient and time shift correction amount of the corresponding sensor channel are obtained by fitting through the synchronization reference point. Then, the local timestamp of each sensor data channel is linearly corrected. After the time correction is completed, a unified time axis is established. For any sensor channel, if there is no corresponding sampling point at any moment on the unified time axis, linear interpolation is performed between two adjacent corrected sampling points to obtain the alignment value of the sensor channel on the unified time axis. Then, the sliding window of a preset size is moved step by step, and the mean and standard deviation within each step of the sliding window are calculated. Based on the mean of each step, the corresponding sensor data of each sensor channel is centered. Then, the standard deviation is used to unify the scale to obtain the standardization result. Finally, the standardized values ​​of the sensor data of all sensor channels at the same moment are concatenated column by column to form a unified observation vector.

[0030] The analysis and extraction module is used to demodulate and perform frequency domain analysis on the swept frequency response signal in multi-source sensor data, and extract the sound wave propagation characteristics.

[0031] Specifically, the analytical extraction module transmits a preset swept-frequency coded signal through the acoustic array. The receiver then segments and samples the original echoes from each receiving channel to establish a discrete signal sequence, simultaneously constructing a corresponding reference swept-frequency excitation. The original echoes and the reference swept-frequency signal are then conjugate-mixed to obtain the demodulated intermediate frequency (IF) or baseband component. After smoothing or low-pass filtering the demodulated IF or baseband component, a corresponding baseband response is generated. Based on the smoothed baseband responses, corresponding analytical signals are constructed. The instantaneous amplitude and phase are obtained from the established analytical signals. The cross-correlation function between each receiving channel and its reference swept-frequency excitation is then calculated. Based on the cross-correlation function, the correlation values ​​at different times are calculated, and the time shift corresponding to the correlation peak is extracted. Finally, the current sampling of the acoustic array is considered. The system calculates the acoustic flight time for each receiving channel, then performs a Discrete Fourier Transform on the smoothed baseband response, extracts the spectral amplitude from the transformed baseband response, and calculates the propagation attenuation at the current frequency based on the known transmission reference spectrum. If the attenuation is linearly related to the propagation distance in any frequency band, the corresponding average propagation attenuation is calculated in the current frequency band to obtain the comprehensive attenuation coefficient index of the corresponding receiving channel. The phase spectrum is then extracted from the transformed baseband response of each receiving channel, and the extracted phase spectrum is expanded to obtain a continuous phase curve. Subsequently, the phase offset relative to the reference spectrum is calculated, and the average phase offset in each frequency band is calculated. Finally, the acoustic flight time, propagation attenuation, and phase change extracted from each receiving channel are integrated into acoustic propagation characteristics.

[0032] The flow field reconstruction module reconstructs the three-dimensional velocity distribution field and turbulence intensity distribution of the protective atmosphere inside the furnace based on the characteristics of sound wave propagation.

[0033] Specifically, based on the furnace's geometric dimensions, inner wall boundaries, and the placement of the acoustic array, a three-dimensional computational domain to be reconstructed is constructed. The furnace's internal space is then discretized into multiple sets of non-overlapping voxel elements. Subsequently, between the acoustic transceiver arrays placed on either side or circumferentially within the furnace, any pair of transmitters and receivers forms an acoustic propagation path. The reciprocal of the propagation speed of each path is calculated and used as the slowness of the corresponding voxel. The equivalent length of each propagation path passing through each voxel element is calculated, and the measured propagation delay of each path is determined. The propagation delay corresponding to each path is then used as an observation vector, and the slowness of all voxel elements is combined to form a parameter vector to be estimated, generating a unified linear inversion model. With the goal of minimizing the sound wave propagation path delay error, a regularization method is used to invert the slowness field of each voxel element. Simultaneously, a stable three-dimensional slowness distribution is obtained by jointly solving the least squares and smoothing constraints. The velocity value in the corresponding voxel element is obtained by calculating the reciprocal of the slowness of each voxel element, and a three-dimensional velocity distribution is constructed using the velocity values ​​of each voxel element. The fluctuation amplitude of the velocity of each voxel element is statistically analyzed at multiple consecutive reconstruction times, and the relative fluctuation amount is used as the turbulence intensity of the corresponding voxel element. Then, the spatial coordinates, velocity values ​​and turbulence intensity values ​​of each voxel are bound to form a complete three-dimensional flow field reconstruction result. Through mapping, the velocity level, flow direction change and turbulence concentration area are displayed in the three-dimensional flow field.

[0034] Example 2, refer to Figure 1 An integrated monitoring system for the coordinated control of protective atmosphere and dew point in a heating furnace includes a data acquisition module, an alignment processing module, an analysis and extraction module, a flow field reconstruction module, an identification and positioning module, a state fusion module, a virtual reconstruction module, an evaluation and judgment module, an exploration and decision-making module, an instruction generation module, and a feedback optimization module.

[0035] The identification and localization module is used to perform time-series evolution analysis on the reconstructed flow field, identify various abnormal modes, and locate potential leakage points or disturbance sources.

[0036] Specifically, the reconstructed 3D flow field results at different times are stored in chronological order, and the local state variables corresponding to anomaly identification are extracted from each spatial voxel unit to form a time-series analysis object. Then, the local baseline of the state characteristics at each time step in each voxel unit is calculated. Based on the acquired local baselines at each time step, the comprehensive deviation of each voxel unit is calculated. The comprehensive deviation of each voxel unit is checked multiple times to see if it exceeds a preset deviation threshold. If it does, and the number of consecutive exceedances exceeds a preset persistence threshold, then the corresponding voxel unit is considered a candidate anomalous voxel. The number of candidate anomalous voxels is counted, and a set of candidate anomalous voxels is constructed. An adjacency graph is established for the candidate anomalous voxels according to their spatial location, and the candidate anomalous voxels are then analyzed. The system expands connectivity to form multiple anomaly clusters, calculates the average anomaly intensity of each cluster, then calculates the center coordinates of each cluster using weighted averages, and calculates the spatial expansion radius based on the center coordinates, average anomaly intensity, and overall deviation of each cluster. The system also obtains the influence range of each cluster, maps the center coordinates, influence range, and average anomaly intensity back to the 3D flow field, generates marker information at the corresponding spatial locations, and calculates the alarm score for each cluster. If the alarm score is higher than the preset comprehensive alarm index, and the average number of continuous exceedances of voxel units within the anomaly cluster is higher than the preset continuous upper limit, an early warning is triggered, the corresponding spatial location is highlighted in red or brightly displayed in the 3D flow field, and the warning information is output.

[0037] The state fusion module constructs a unified state vector and incorporates historical control behavior sequences to form a high-dimensional state vector; the virtual reconstruction module generates dew point data in real time based on various sensor data.

[0038] Specifically, when any dew point sensor malfunctions or fails, an abnormality score is calculated based on the rolling median baseline formed by historical effective dew point values ​​and a reference dew point value. If the abnormality score exceeds a preset threshold, the dew point data at that moment is determined to be abnormal or failed, and that time period is marked as a compensation interval. Simultaneously, the output of that dew point sensor is stopped, and synchronous data from the other corresponding sensors at the same time is extracted. This includes in-furnace infrared spectral characteristics, oxygen potential signal, furnace temperature distribution, furnace pressure fluctuation, and historical effective dew point sequences. Each synchronous data point is feature-encoded, and a random noise vector is added to construct a joint input vector. This joint input vector is then fed into a condition generator. The generator outputs candidate dew point data for the current missing time or missing interval point by point according to the time window. Then, the candidate dew point data is sent to the discriminator for authenticity evaluation and a credibility score is generated for each candidate dew point data. If the credibility score is higher than the preset threshold, it indicates that the corresponding candidate dew point data is acceptable. At the same time, the allowable interval of dew point data and the maximum allowable adjacent variation are set, and the physical constraint penalty term for each dew point data at the current time is calculated. If the physical constraint penalty term is higher than the preset intervention threshold, the candidate dew point data at the current time is compressed and corrected. After completing the discrimination and physical correction, the candidate dew point data is weighted and fused according to the credibility score and the physical penalty value to generate the final compensated dew point value.

[0039] The evaluation and judgment module is used to dynamically evaluate the fused high-dimensional state vector and identify each critical state; the exploration and decision module uses the current state and historical feedback to learn and optimize the control strategy.

[0040] Specifically, at each decision-making moment, the current operating state and historical feedback information are collected and organized into a control state vector. Based on the current control state vector, the initial control action is output according to the current policy network, and random exploration noise is added to perturb the initial control action. The execution action generated by the perturbation is sent to the controller. After the controller executes the action, the immediate reward is calculated based on the actual response. Then, the current state, action, reward, and next state are combined into a time series sample, and the target Q value is calculated based on the current time series sample. Subsequently, the time series difference loss is calculated, and historical interaction samples are stored in the experience pool in sequence. Multiple samples are extracted from the experience pool, and the corresponding batch average loss is calculated based on the time series difference loss of each sample. If the batch average loss is higher than the preset loss value, the policy network parameters are updated, and the execution action is generated again. The process is iteratively updated until the batch average loss is lower than the preset loss value. Through training, the statistical policy network generates control policies at different times and calculates their average loss and average reward. If the average loss is lower than the preset threshold and the preset stability condition is met, the policy corresponding to the current parameters is output as the optimal control policy.

[0041] The instruction generation module generates multi-variable collaborative control instructions based on the control strategy; the feedback optimization module collects the system response data after execution in real time, compares it with the prediction results, and dynamically corrects the control strategy.

[0042] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. An integrated monitoring system for the coordinated control of protective atmosphere and dew point in a heating furnace, characterized in that, The system includes a data acquisition module, an alignment processing module, an analytical extraction module, a flow field reconstruction module, an identification and positioning module, a state fusion module, a virtual reconstruction module, an evaluation and judgment module, an exploration and decision-making module, an instruction generation module, and a feedback optimization module. The data acquisition module collects real-time sensor data on the propagation characteristics, thermal state, and composition information of the furnace atmosphere from various sources. The alignment processing module aligns the time of the multi-source sensor data and constructs a standardized observation vector. The analytical extraction module demodulates and performs frequency domain analysis on the swept-frequency response signal from the multi-source sensor data and extracts the sound wave propagation characteristics. The flow field reconstruction module reconstructs the three-dimensional velocity distribution field and turbulence intensity distribution of the protective atmosphere inside the furnace based on the sound wave propagation characteristics. The identification and localization module performs time-series evolution analysis on the reconstructed flow field, identifies various abnormal patterns, and locates potential leakage points or disturbance sources. The state fusion module constructs a unified state vector and incorporates historical control behavior sequences to form a high-dimensional state vector. The virtual reconstruction module generates dew point data in real time based on various sensor data. The evaluation and judgment module dynamically evaluates the fused high-dimensional state vector and identifies various critical states. The exploration and decision-making module learns and optimizes the control strategy using the current state and historical feedback. The instruction generation module generates multi-variable collaborative control instructions according to the control strategy. The feedback optimization module collects the system response data after execution in real time and compares it with the prediction results to dynamically correct the control strategy.

2. The integrated monitoring system for coordinated control of protective atmosphere and dew point in a heating furnace according to claim 1, characterized in that, The specific steps of the alignment processing module in aligning multi-source sensor data time and constructing standardized observation vectors are as follows: S1.1: Retain the original sampled value and local timestamp of each channel's sensor data, and bind the original sampled value and local timestamp of each channel's sensor data to form the time series of each channel. Then, collect the original time of each sensor channel at two synchronization moments, and obtain the clock scale correction coefficient and time shift correction amount of the corresponding sensor channel by fitting the synchronization reference point. Then, perform linear correction on the local timestamp of each channel's sensor data. S1.2: After completing the time correction, a unified time axis is established. For any sensing channel, if there is no corresponding sampling point at any time on the unified time axis, linear interpolation is performed between two adjacent corrected sampling points to obtain the alignment value of the sensing channel on the unified time axis. Then, the sliding window of a preset size is moved step by step, and the mean and standard deviation within each sliding window are calculated. S1.3: Based on the mean of each step, the corresponding sensor data of each sensor channel is centered, and then the standard deviation is used to unify the scale to obtain the standardized result. Finally, the standardized values ​​of the sensor data of all sensor channels at the same time are concatenated column by column to form a unified observation vector.

3. The integrated monitoring system for coordinated control of protective atmosphere and dew point in a heating furnace according to claim 2, characterized in that, The detailed steps of the analytical extraction module for demodulating and performing frequency domain analysis on the swept frequency response signal in the multi-source sensor data are as follows: S2.1: The parsing and extraction module transmits a preset sweep frequency encoded signal through the acoustic wave array. Then, the receiver segments and samples the original echoes of each receiving channel to establish a discrete signal sequence and construct a corresponding reference sweep frequency excitation. After that, the original echoes and the reference sweep frequency signal are conjugate mixed to obtain the demodulated intermediate frequency or baseband components. S2.2: After performing sliding smoothing or low-pass filtering on the demodulated intermediate frequency or baseband components, the corresponding baseband response is generated. Based on the smoothed baseband responses, the corresponding analytical signals are constructed. Based on the established analytical signals, the corresponding instantaneous amplitude and instantaneous phase are obtained. Then, the cross-correlation function between each receiving channel and its reference sweep frequency excitation is calculated. S2.3: Based on the cross-correlation function, calculate the correlation value at different times, extract the time shift corresponding to the correlation peak, and calculate the flight time of the acoustic wave in each receiving channel according to the current sampling period of the acoustic array. Then, perform a discrete Fourier transform on the smoothed baseband response, extract the spectral amplitude from the transformed baseband response, and calculate the propagation attenuation at the current frequency according to the known transmission reference spectrum. S2.4: If the attenuation is linearly related to the propagation distance in any frequency band, calculate the corresponding average propagation attenuation in the current frequency band to obtain the comprehensive attenuation coefficient index of the corresponding receiving channel. Then, extract the phase spectrum from the baseband response after transformation of each receiving channel, and perform phase expansion on each extracted phase spectrum to obtain a continuous phase curve. Subsequently, calculate the phase offset relative to the reference spectrum and calculate the average phase offset in each frequency band. Then, integrate the sound wave flight time, propagation attenuation and phase change extracted from each receiving channel into sound wave propagation characteristics.

4. The integrated monitoring system for coordinated control of protective atmosphere and dew point in a heating furnace according to claim 3, characterized in that, The specific steps of the flow field reconstruction module to reconstruct the three-dimensional velocity distribution field and turbulence intensity distribution of the protective atmosphere inside the furnace are as follows: S3.1: Based on the geometric dimensions of the furnace, the inner wall boundary, and the layout of the acoustic array, construct the three-dimensional computational domain to be reconstructed. Then, discretize the internal space of the furnace into multiple sets of non-overlapping voxel units. Then, between the acoustic transceiver arrays arranged on both sides or around the furnace, any pair of transmitters and receivers constitutes an acoustic propagation path. Calculate the reciprocal of the propagation speed of each acoustic propagation path and use it as the slowness of the corresponding voxel. S3.2: Calculate the equivalent length of each propagation path through each voxel unit, and calculate the measured propagation delay of each sound wave propagation path. Then, use the propagation delay corresponding to each sound wave propagation path as the observation vector, and combine the slowness of all voxel units into the parameter vector to be estimated, so as to generate a unified linear inversion model. S3.3: With the goal of minimizing the time delay error of each sound wave propagation path, a regularization method is used to invert the slowness field of each voxel unit. At the same time, a stable three-dimensional slowness distribution is obtained by jointly solving the least squares and smoothing constraints. The velocity value in the corresponding voxel unit is obtained by calculating the reciprocal of the slowness of each voxel unit, and a three-dimensional velocity distribution is constructed by the velocity values ​​of each voxel unit. S3.4: Statistically analyze the fluctuation range of the velocity of each voxel unit at multiple consecutive reconstruction times, and use the relative fluctuation amount as the turbulence intensity of the corresponding voxel unit. Then, bind the spatial coordinates, velocity value and turbulence intensity value of each voxel to form a complete three-dimensional flow field reconstruction result. Through mapping, display the velocity level, flow direction change and turbulence concentration area in the three-dimensional flow field.

5. The integrated monitoring system for coordinated control of protective atmosphere and dew point in a heating furnace according to claim 1, characterized in that, The specific steps of the identification and positioning module in identifying various abnormal patterns and locating potential leakage points or disturbance sources are as follows: S4.1: Store the three-dimensional flow field results obtained from reconstruction at different times in chronological order, and extract the local state variables corresponding to the anomaly identification in each spatial voxel unit to form a time-series analysis object. Then calculate the local baseline of the state characteristics of each voxel unit at each time, and calculate the comprehensive deviation of each voxel unit based on the obtained local baseline at each time. S4.2: Detect whether the overall deviation of each voxel unit is greater than the preset deviation threshold multiple times. If it is greater and the number of consecutive exceedances is higher than the preset persistence judgment threshold, it indicates that the corresponding voxel unit is a candidate abnormal voxel. Count the number of candidate abnormal voxels and construct a set of candidate abnormal voxels. S4.3: Establish an adjacency graph for candidate anomalous voxels according to their spatial location, and expand the candidate anomalous voxels to form multiple anomalous clusters. At the same time, calculate the average anomalous intensity of each anomalous cluster, then calculate the center coordinates of each anomalous cluster by weighting, and calculate its spatial expansion radius based on the center coordinates, average anomalous intensity and comprehensive deviation of each anomalous cluster, and obtain the influence range of each anomalous cluster. S4.4: Map the center coordinates, influence range and average anomaly intensity of the anomaly cluster back to the three-dimensional flow field, generate marker information at the corresponding spatial location, and calculate the alarm score of each anomaly cluster. If the alarm score is higher than the preset comprehensive alarm index and the average number of continuous over-limits of the voxel units in the anomaly cluster is higher than the preset continuous upper limit, an early warning is triggered, and the corresponding spatial location is marked in red or highlighted in the three-dimensional flow field, and the early warning information is output.

6. The integrated monitoring system for coordinated control of protective atmosphere and dew point in a heating furnace according to claim 1, characterized in that, The specific steps by which the virtual reconstruction module generates dew point data in real time based on data from various sensors are as follows: S5.1: When any dew point sensor malfunctions or fails, the corresponding dew point sensor malfunction score is calculated based on the rolling median baseline formed by the historical effective dew point values ​​and the reference dew point value. If the malfunction score is higher than the preset threshold, the dew point data at that moment is determined to be abnormal or failed, and that time period is marked as the interval to be compensated. At the same time, the output of the dew point sensor is stopped, and synchronous data at the same moment is extracted from the other corresponding sensors, including the furnace infrared spectral characteristics, oxygen potential signal, furnace temperature distribution, furnace pressure fluctuation, and historical effective dew point sequence. S5.2: Encode the features of each synchronization data, add a random noise vector, and construct a joint input vector. Then, send the joint input vector to the condition generator. The condition generator outputs the candidate dew point data of the current missing time or missing interval point by point according to the time window. Then, send the candidate dew point data to the discriminator for authenticity evaluation and generate a confidence score for each candidate dew point data. S5.3: If the credibility score is higher than the preset threshold, it indicates that the corresponding candidate dew point data is acceptable. At the same time, the allowable range of dew point data and the maximum allowable adjacent variation are set, and the physical constraint penalty term of each dew point data at the current time is calculated. If the physical constraint penalty term is higher than the preset intervention threshold, the candidate dew point data at the current time is compressed and corrected. After completing the discrimination and physical correction, the candidate dew point data is weighted and fused according to the credibility score and the physical penalty value to generate the final compensated dew point value.

7. The integrated monitoring system for coordinated control of protective atmosphere and dew point in a heating furnace according to claim 1, characterized in that, The steps of the exploration decision module in learning and optimizing the control strategy are as follows: S6.1: At each decision moment, collect the current operating state and historical feedback information, organize them into a control state vector, and output the initial control action according to the current policy network based on the current control state vector, and add random exploration noise to perturb the initial control action; S6.2: Send the actions generated by the disturbance to the controller, and calculate the immediate reward based on the actual response after the controller executes the action. Then, combine the current state, action, reward and the next state into a time series sample, calculate the target Q value based on the current time series sample, calculate the time series difference loss, and store the historical interaction samples into the experience pool in order. S6.3: Extract multiple samples from the experience pool and calculate the corresponding batch average loss based on the temporal difference loss of each sample. If the batch average loss is higher than the preset loss value, update the policy network parameters and generate the execution action again. Continue to iterate and update until the batch average loss is lower than the preset loss value. S6.4: The statistical policy network, after training, generates control policies at different times and calculates their average loss and average reward. If the average loss is lower than the preset threshold and the preset stability condition is met, the policy corresponding to the current parameters is output as the optimal control policy.

8. The integrated monitoring system for coordinated control of protective atmosphere and dew point in a heating furnace according to claim 1, characterized in that, The control state vector described in S6.1 specifically includes the target deviation of the current controlled object, the current atmosphere state deviation, the current pressure state deviation, the current temperature state deviation, and a summary of historical control feedback.

Images