Combustion refinement adjustment method based on whole furnace expansion monitoring

Abstract

The present application relates to the technical field of mechanical stress and thermal deformation monitoring, and particularly relates to a combustion refinement adjustment method based on full-furnace expansion monitoring, wherein an expansion sensor array is arranged at each layer in the main burner area of a boiler furnace, a vibration monitoring unit is synchronously arranged in the layering from the main burner to the SOFA air area, normal deformation amount, three-dimensional expansion displacement data and vibration frequency spectrum data of the four furnace walls are collected; standardized monitoring data sets are generated through wavelet denoising and temperature compensation processing, a reinforcement learning model is input by fusing boiler operation parameters to generate a burner adjustment instruction; model weight is optimized by real-time collection of water cooling wall expansion rate change data, a control strategy is updated, and a closed-loop feedback mechanism is formed. Through multi-source sensing data fusion and dynamic modeling, sub-millimeter level analysis of thermal load distribution and rapid response of combustion state are realized, and the operation efficiency of the boiler and the service life of the equipment are improved.

Description

Technical Field

[0001] This invention relates to the field of mechanical stress and thermal deformation monitoring technology, and in particular to a method for refined combustion adjustment based on full furnace expansion monitoring. Background Technology

[0002] Existing boiler combustion control technologies generally employ water-cooled wall outlet temperature monitoring or two-dimensional temperature field analysis of the combustion zone. These methods suffer from limitations such as single monitoring dimension and insufficient spatial coverage, making it difficult to capture real-time dynamic heat load distribution and local flame deviation. Particularly in spiral water-cooled wall boilers, the coupling effect of medium flow rate and heat load easily leads to temperature monitoring data being affected by non-uniform heat transfer, failing to accurately characterize the actual heat flux density of the heating surface. Furthermore, under low-load peak-shaving conditions, flame deflection and imbalanced thermal stress distribution in the furnace are exacerbated. Existing methods lack direct correlation analysis of water-cooled wall deformation characteristics, making it difficult to dynamically identify local overload areas and combustion instability risks, resulting in delayed combustion adjustment, thermal efficiency loss, and reduced equipment lifespan. Therefore, a technical solution integrating thermal deformation monitoring and dynamic combustion control is urgently needed to overcome the shortcomings of existing methods in terms of spatial resolution, real-time performance, and adaptability to operating conditions. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for refined combustion adjustment based on full furnace expansion monitoring, which solves the problem of lag in combustion adjustment caused by insufficient spatial resolution in existing temperature monitoring methods.

[0004] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0005] The present invention provides a method for fine-tuning combustion based on full-furnace expansion monitoring, comprising:

[0006] An array of expansion sensors is arranged in each layer of the main burner area of ​​the boiler furnace to collect the normal deformation and three-dimensional expansion displacement data of the four furnace walls. Vibration monitoring units are set up in layers from the main burner area to the SOFA wind area to obtain vibration spectrum data.

[0007] The normal deformation, three-dimensional expansion displacement data, and vibration spectrum data are subjected to wavelet noise reduction filtering and temperature compensation processing respectively to generate a standardized monitoring dataset including displacement features and vibration features.

[0008] Based on the three-dimensional expansion displacement of the four furnace walls at the same height in the standardized monitoring dataset, the expansion deviation rate of the symmetrical monitoring points is calculated. Based on the expansion deviation rate, a flame center offset vector model is constructed using the random forest algorithm, and a heat load distribution cloud map reflecting the spatial distribution of heat load is output.

[0009] The coordinates of monitoring points whose normal deformation exceeds the material's elastic limit are extracted from the standardized monitoring dataset, and the energy proportion characteristics of the low-frequency band in the vibration spectrum data of the corresponding location are associated with them. Local areas with abnormal heat flux density are identified by a support vector machine classifier.

[0010] The heat load distribution cloud map output by the flame center offset vector model, the coordinate information of the abnormal area identified by the support vector machine classifier, and the oxygen content, NOx concentration, furnace outlet temperature, desuperheating water volume and steam temperature data in the boiler operating parameters are input into the reinforcement learning model to generate burner output adjustment commands and damper opening combination parameters.

[0011] The distributed control system sends the burner output adjustment command and damper opening combination parameters to the burner actuator.

[0012] Real-time data on the expansion rate change of the water-cooled wall within a preset time period after adjustment is collected. Based on the expansion rate change data, the weight parameters of the reinforcement learning model are optimized using the gradient descent method, and the control strategy in the combustion adjustment rule base is updated.

[0013] Furthermore, the method for refined combustion adjustment based on full furnace expansion monitoring, wherein the expansion sensor array is arranged in each layer of the main burner area of ​​the boiler furnace to collect the normal deformation and three-dimensional expansion displacement data of the four furnace walls, and vibration monitoring units are set up in layers from the main burner area to the SOFA wind area to obtain vibration spectrum data, includes: setting up 4 sets of sensor nodes in each height layer of the four main burner areas of the furnace, each set of nodes including a normal displacement sensor and a three-dimensional dilatometer, for collecting the normal deformation and three-dimensional expansion displacement data of the four furnace walls;

[0014] The vibration monitoring unit is arranged in layers from the main burner area to the SOFA wind area. Each layer is equipped with a vibration acceleration sensor and a spectrum analysis module to obtain vibration spectrum data of the corresponding area.

[0015] Furthermore, the method for refined combustion adjustment based on full furnace expansion monitoring, wherein the normal deformation, three-dimensional expansion displacement data and vibration spectrum data are respectively processed by wavelet noise reduction filtering and temperature compensation to generate a standardized monitoring dataset including displacement features and vibration features, includes: using a wavelet transform algorithm to filter out high-frequency noise from the original vibration spectrum data collected by the vibration monitoring unit;

[0016] The temperature compensation process includes: linearly correcting the expansion displacement measurement value obtained by the expansion sensor array based on the furnace ambient temperature data to eliminate the deviation of the thermal expansion coefficient.

[0017] Furthermore, the combustion fine-tuning method based on full furnace expansion monitoring involves calculating the expansion deviation rate of symmetrical monitoring points based on the three-dimensional expansion displacement of the four furnace walls at the same height level in the standardized monitoring dataset, constructing a flame center offset vector model using a random forest algorithm based on the expansion deviation rate, and outputting a heat load distribution cloud map reflecting the spatial distribution of heat load. The expansion deviation rate of symmetrical monitoring points at the same height level is also calculated based on the expansion displacement of the four furnace walls in the standardized monitoring dataset.

[0018] The mapping relationship between the expansion deviation rate and the flame center coordinates is established by using the random forest algorithm, and a heat load spatial distribution matrix with weight coefficients is generated.

[0019] Furthermore, the refined combustion adjustment method based on full furnace expansion monitoring, which involves extracting the coordinates of monitoring points whose normal deformation exceeds the material's elastic limit from the standardized monitoring dataset, associating them with the energy proportion characteristics of the low-frequency band in the vibration spectrum data at the corresponding locations, and identifying local heat flux density anomaly regions using a support vector machine classifier, includes:

[0020] The coordinates of monitoring points whose normal deformation exceeds the elastic limit of the water-cooled wall tube are extracted from the standardized monitoring dataset. The elastic limit is dynamically set based on the yield strength threshold of the water-cooled wall tube and in combination with the real-time monitored furnace temperature.

[0021] By combining the temporal energy distribution characteristics in the vibration spectrum data at the corresponding location, a support vector machine classifier is used to determine the risk level of combustion instability.

[0022] A set of anomaly region coordinates, including priority labels, is generated and used as input parameters for the reinforcement learning model.

[0023] Furthermore, in the aforementioned method for refined combustion adjustment based on full furnace expansion monitoring, the workflow of the reinforcement learning model includes:

[0024] The generated spatial distribution matrix of heat load is spatially encoded with the set of coordinates of abnormal areas;

[0025] The data include NOx concentration, oxygen content, furnace outlet temperature, desuperheating water volume, and steam temperature from the boiler operating parameters.

[0026] The combination of adjustment parameters under the multi-objective constraints of heat load balancing, NOx emission minimization, and steam temperature stability is solved by a near-end strategy optimization algorithm.

[0027] Furthermore, the combustion fine-tuning method based on full furnace expansion monitoring, wherein the real-time acquisition of water-cooled wall expansion rate change data within a preset time period after adjustment, the optimization of the weight parameters of the reinforcement learning model based on the expansion rate change data using gradient descent, and the updating of the control strategy in the combustion adjustment rule base include:

[0028] Collect data on the change in the water-cooled wall expansion rate within a preset time range after the burner actuator completes its adjustment action;

[0029] Calculate the expansion suppression efficiency coefficient of the current adjustment strategy based on the expansion rate change data;

[0030] Based on the inflation suppression efficiency coefficient, the weight parameters of the reinforcement learning model are optimized and the reward function value is updated using the gradient descent method.

[0031] Furthermore, the method for fine-tuning combustion based on full-furnace expansion monitoring also includes:

[0032] The expansion displacement pattern, vibration spectrum characteristics, and combustion parameters from historical operating data are combined and input into a convolutional neural network.

[0033] Spatiotemporal correlation feature vectors are extracted and dimensionality reduced by principal component analysis to generate a combustion health index.

[0034] The combustion state is classified into three categories based on the combustion health index: steady state, wave dynamics, and unstable state, and the classification results are fed back to the reinforcement learning model.

[0035] Furthermore, the method for fine-tuning combustion based on full-furnace expansion monitoring also includes:

[0036] Based on coal quality change data, a transfer learning algorithm is invoked to adapt the parameters of the reinforcement learning model.

[0037] The kernel function parameters of the support vector machine classifier are dynamically adjusted based on the classification results of the combustion health index.

[0038] The combustion health index determination threshold of the convolutional neural network is updated based on the range of changes in coal calorific value.

[0039] Furthermore, the combustion fine adjustment method based on full furnace expansion monitoring also includes: after each combustion adjustment command is executed, collecting expansion displacement correlation data of two adjacent monitoring points arranged at preset intervals along the circumference of the furnace in the same height layer, and four monitoring points in the upper and lower adjacent layers.

[0040] Based on the aforementioned correlation data, a transfer function relating vibration spectrum characteristics to changes in heat flux density is constructed.

[0041] The heat load weighting coefficient of the flame center offset vector model is corrected by the transfer function.

[0042] Furthermore, in the combustion fine-tuning method based on full furnace expansion monitoring described in this invention, the expansion sensor array is arranged as follows: a wire-based distance measuring hook, a wire-based distance measuring sensor, and a wire-based distance measuring fixture. The wire-based distance measuring hook is fixed to the surface of the furnace outer wall insulation sheet at preset intervals. The wire-based distance measuring sensor is connected to the hook via a wire to form a distance measuring grid (e.g., ...). Figure 4 ), used to collect the normal deformation and three-dimensional expansion displacement data of the four furnace walls in real time;

[0043] The vibration monitoring unit includes a vibration measurement module and a vibration measurement probe. The vibration measurement probe is installed on the surface of the furnace water-cooled wall through a fixed bracket of the full furnace expansion monitoring device to obtain vibration spectrum data of the corresponding area.

[0044] The computing and control host module receives real-time data from the wire distance sensor, vibration measurement probe, and wall temperature measurement probe via a communication antenna, and generates combustion optimization instructions based on load conditions, coal quality parameters, and air volume data.

[0045] Beneficial effects of this invention;

[0046] The beneficial effects of this invention are as follows: by fusing multi-source data from a full-furnace expansion sensor array and a vibration monitoring unit, and combining wavelet denoising and temperature compensation processing, a high-precision standardized monitoring dataset is generated. A random forest algorithm is used to construct a flame center offset vector model to analyze the sub-meter-level heat load distribution. Simultaneously, a support vector machine classifier is used to associate deformation exceeding limits and vibration spectrum features to achieve dynamic identification of local abnormal areas. Combined with a reinforcement learning model, combustion adjustment commands are generated under multi-objective constraints, and a closed-loop feedback optimization mechanism is formed. This effectively improves the spatial resolution of heat load and the response speed of combustion status, reduces the risk of misjudgment of flame offset caused by a single monitoring dimension, and enhances boiler thermal efficiency and equipment operation safety. Attached Figure Description

[0047] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0048] Figure 1 A flowchart of a combustion fine adjustment method based on full furnace expansion monitoring provided in an embodiment of the present invention.

[0049] Figure 2 This is a cross-sectional schematic diagram of the entire furnace from a frontal perspective, provided for an embodiment of the present invention.

[0050] Figure 3 This is a top view of the entire furnace provided in an embodiment of the present invention.

[0051] Figure 4 This is a schematic diagram of the grid arrangement connection method for the first measuring device provided in an embodiment of the present invention.

[0052] Figure 5 This is a schematic diagram of the connection method of the wire ranging grid arrangement of the second measuring device provided in an embodiment of the present invention.

[0053] Figure 6 This is a schematic diagram of the layered vibration monitoring units from the main burner area to the SOFA wind area provided in an embodiment of the present invention.

[0054] Explanation of reference numerals in the attached drawings: 1. Wire distance measuring hook; 2. Wire distance measuring sensor; 3. Wire distance measuring fixture; 4. Indicator light for the whole furnace expansion monitoring device; 5. Power light for the whole furnace expansion monitoring device; 6. Communication antenna for the whole furnace expansion monitoring device; 7. Calculation and control host module for the whole furnace expansion monitoring device; 8. Vibration measurement module; 9. Wall temperature measurement module; 10. Wall temperature measurement probe; 11. Vibration measurement probe; 12. Boiler furnace water-cooled wall; 13. Fixing bracket for the whole furnace expansion monitoring device; 14. Housing for the whole furnace expansion monitoring device; 15. Fixing hole for the whole furnace expansion monitoring device; 16. Sheet metal for furnace outer wall insulation; 17. Furnace outer wall insulation layer. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings. To better understand the objectives of this invention, it will be described in further detail below.

[0056] Please see Figures 1 to 6 The present invention provides a method for fine-tuning combustion based on full-furnace expansion monitoring, comprising:

[0057] Step S101: An array of expansion sensors is arranged in each layer of the main burner area of ​​the boiler furnace to collect the normal deformation and three-dimensional expansion displacement data of the four furnace walls, and vibration monitoring units are set in layers from the main burner area to the SOFA wind area to obtain vibration spectrum data.

[0058] An expansion sensor array is arranged in each layer of the main burner area of ​​the boiler furnace. The array includes a wire-based distance measuring hook 1, a wire-based distance measuring sensor 2, and a wire-based distance measuring fixture 3. The wire-based distance measuring hook 1 is fixed to the surface of the furnace outer wall insulation sheet 16 at preset intervals. The wire-based distance measuring sensor 2 is connected to the hook 1 via a wire to form a distance measuring grid. Figure 4 It is used to collect the normal deformation and three-dimensional expansion displacement data of the furnace wall in real time.

[0059] Vibration monitoring units are synchronously set up in layers from the main burner to the SOFA air area. Each unit includes a vibration measurement module 8 and a vibration measurement probe 11. The vibration measurement probe 11 is installed on the surface of the furnace water-cooled wall 12 through the fixing bracket 13 of the whole furnace expansion monitoring device to obtain vibration spectrum data of the corresponding area.

[0060] Furthermore, a wall temperature measurement module 9 and a wall temperature measurement probe 10 are arranged on the surface of the furnace water-cooled wall 12 to collect water-cooled wall temperature data in real time.

[0061] The whole furnace expansion monitoring and control host module 7 receives real-time data from the wire distance sensor 2, vibration measurement probe 11 and wall temperature measurement probe 10 through the communication antenna 6, and summarizes the expansion deformation data, vibration data and wall temperature data into the whole furnace expansion monitoring and control host according to type.

[0062] The main unit of the computing and control system classifies and processes data based on the current load conditions, coal quality parameters, primary and secondary air volume and temperature, identifies abnormal values ​​through real-time calculation and generates combustion optimization operation instructions.

[0063] Step S102: Wavelet noise reduction filtering and temperature compensation processing are performed on the normal deformation, three-dimensional expansion displacement data and vibration spectrum data respectively to generate a standardized monitoring dataset including displacement features and vibration features.

[0064] Step S103: Based on the three-dimensional expansion displacement of the four furnace walls at the same height in the standardized monitoring dataset, calculate the expansion deviation rate of the symmetrical monitoring points, and construct the flame center offset vector model using the random forest algorithm based on the expansion deviation rate to output a heat load distribution cloud map reflecting the spatial distribution of heat load.

[0065] Step S104: Extract the coordinates of monitoring points whose normal deformation exceeds the elastic limit of the material from the standardized monitoring dataset, and associate them with the energy proportion features of the low-frequency band in the vibration spectrum data of the corresponding location. Identify the local heat flux density abnormal area through the support vector machine classifier.

[0066] Step S105: Input the heat load distribution cloud map output by the flame center offset vector model, the coordinate information of the abnormal area identified by the support vector machine classifier, and the oxygen content, NOx concentration, furnace outlet temperature, desuperheating water volume and steam temperature data in the boiler operating parameters into the reinforcement learning model to generate burner output adjustment command and damper opening combination parameters.

[0067] Step S106: The burner output adjustment command and damper opening combination parameters are sent to the burner actuator through the distributed control system.

[0068] Step S107: Real-time acquisition of water-cooled wall expansion rate change data within a preset time period after adjustment; optimization of the weight parameters of the reinforcement learning model based on the expansion rate change data using gradient descent method; and updating of the control strategy in the combustion adjustment rule base.

[0069] An expansion sensor array is installed in each layer of the main burner area of ​​the boiler furnace. The array is distributed across each height layer of the four furnace walls. Each sensor node includes a normal displacement sensor and a three-dimensional dilatometer, used to collect real-time data on the normal deformation and three-dimensional expansion displacement of the furnace walls. Vibration monitoring units are arranged layer by layer from the main burner area to the SOFA (Solar Air Flow) area. Each layer is equipped with a vibration acceleration sensor and a spectrum analysis module to synchronously acquire vibration spectrum data for the corresponding area. Through this multi-sensor collaborative layout, a monitoring network for the mechanical deformation and vibration status of the entire furnace area is constructed.

[0070] The collected raw data requires signal preprocessing. For vibration spectrum data, wavelet transform algorithm is used to separate high-frequency noise from effective signals, filtering out irrelevant noise components introduced by equipment vibration or electromagnetic interference. For expansion sensor data, based on real-time monitoring values ​​of furnace ambient temperature, linear correction is applied to the expansion displacement measurements to eliminate measurement errors caused by differences between the material's thermal expansion coefficient and ambient temperature, generating standardized displacement and vibration feature datasets.

[0071] Based on a standardized dataset, the three-dimensional expansion displacement deviation rate of symmetrical monitoring points on four furnace walls at the same height level is calculated. A random forest algorithm is used to extract features and perform pattern matching on multiple sets of expansion deviation rate data, establishing a mapping relationship between the expansion deviation rate and the spatial position of the flame center, generating a heat load distribution cloud map reflecting the spatial distribution of heat load. This model, through multi-dimensional data fusion, analyzes the sub-meter-level distribution characteristics of the heat load inside the furnace.

[0072] Based on displacement feature data, the coordinates of monitoring points where the normal deformation exceeds the elastic limit of the water-cooled wall pipe are identified. Combining the energy distribution characteristics of the low-frequency band in the vibration spectrum at the corresponding location, a support vector machine classifier is used to analyze the correlation between the time-domain characteristics of the vibration signal and the heat flux density anomaly, determine the risk level of combustion instability in the local area, and generate a set of anomaly area coordinates with priority labels.

[0073] The heat load distribution cloud map, the coordinate set of abnormal areas, and boiler operating parameters such as oxygen content, NOx concentration, furnace outlet temperature, desuperheating water flow rate, and steam temperature are input into the reinforcement learning model. Using a near-end strategy optimization algorithm, the optimal parameter combination for burner output adjustment and damper opening is solved under multi-objective constraints. This model integrates thermodynamic balance, pollutant emission control, and equipment safety operation requirements to dynamically generate combustion adjustment commands.

[0074] The adjustment command is sent to the burner actuator through the distributed control system, driving the coordinated adjustment of burner output and damper opening. During execution, real-time data on the expansion rate change of the water-cooled wall is collected to quantify the effect of combustion adjustment on suppressing furnace mechanical deformation. Based on the expansion rate change trend, the gradient descent method is used to optimize the weight parameters of the reinforcement learning model, update the control strategy in the rule base, and form a closed-loop feedback mechanism.

[0075] The above steps, through a closed-loop process of data acquisition, feature analysis, intelligent decision-making, and feedback optimization, achieve dynamic coupling analysis of combustion state and mechanical deformation. Sensor layout and signal preprocessing provide standardized input for model construction; random forest algorithm and support vector machine classifier solve the problems of heat load distribution analysis and abnormal area identification, respectively; reinforcement learning model generates control commands under multi-objective constraints and continuously optimizes the decision logic through execution feedback.

[0076] Specifically, the method for refined combustion adjustment based on full furnace expansion monitoring includes arranging expansion sensor arrays in each layer of the main burner area of ​​the boiler furnace to collect the normal deformation and three-dimensional expansion displacement data of the four furnace walls, and setting vibration monitoring units in layers from the main burner area to the SOFA wind area to obtain vibration spectrum data. This includes setting 4 sets of sensor nodes in each height layer of the four main burner areas of the furnace, with each set of nodes including a normal displacement sensor and a three-dimensional dilatometer, for collecting the normal deformation and three-dimensional expansion displacement data of the four furnace walls.

[0077] The vibration monitoring unit is arranged in layers from the main burner area to the SOFA wind area. Each layer is equipped with a vibration acceleration sensor and a spectrum analysis module to obtain vibration spectrum data of the corresponding area.

[0078] Four sets of sensor nodes are arranged at each height level in the main burner area on all four sides of the furnace. Each set of nodes consists of a normal displacement sensor and a three-dimensional dilatometer, and is installed on a preset mounting reference surface on each of the four furnace walls. The normal displacement sensor is arranged along the normal direction of the furnace wall and coupled to the furnace wall surface through a rigid fixing bracket, and is used to measure the deformation perpendicular to the plane of the furnace wall in real time. The three-dimensional dilatometer adopts a multi-axis displacement sensing structure, and its three measuring axes correspond to the axial, radial and tangential directions in the furnace coordinate system, respectively, to synchronously collect the three-dimensional expansion displacement components of the furnace wall. The spatial layout of the four sets of nodes forms a symmetrical monitoring network of the four furnace walls at the same height level, covering the full circumferential deformation characteristics of the burner area.

[0079] Each sensor node requires initial orientation calibration during installation. The measurement axis of the normal displacement sensor is aligned with the furnace wall normal direction based on the furnace geometry parameters, and a mapping relationship is established between the coordinate system of the 3D dilatometer and the global coordinate system of the furnace. Sensor nodes achieve synchronous sampling through a distributed data acquisition module, eliminating phase errors caused by signal transmission delays. Monitoring data from each height level undergoes redundancy verification to generate a deformation dataset with spatiotemporal consistency.

[0080] The vibration monitoring units are arranged in layers from the main burner area to the SOFA air area. Each layer is equipped with a vibration acceleration sensor array, and its installation position is at a preset angle to the burner nozzle axis to optimize the sensitivity of the vibration signal in the direction of combustion disturbance. The vibration acceleration sensors adopt a wideband response design, covering the characteristic frequency bands of combustion pulsation and mechanical vibration. The spectrum analysis module is integrated into the data processing unit of each monitoring layer, performing time-domain segmentation and frequency-domain windowing processing on the raw vibration signal to generate vibration spectrum data including energy spectrum, kurtosis, and envelope characteristics.

[0081] The layered arrangement strategy of the vibration monitoring units matches the airflow path from the burner area to the SOFA wind area. The main burner area monitoring layer focuses on capturing low-frequency vibration components caused by combustion flame pulsation, while the SOFA wind area monitoring layer targets mid-to-high-frequency vibration characteristics generated by secondary wind disturbances. The spectrum analysis module of each layer extracts vibration mode parameters related to the combustion state through a characteristic frequency band segmentation algorithm, forming a vibration feature set that is spatiotemporally aligned with the expansion monitoring data.

[0082] Sensor nodes and vibration monitoring units achieve data fusion through redundant communication links. The three-dimensional displacement data from the expansion sensor and the spectral characteristic data from the vibration monitoring are timestamped before being input into a multi-source data verification module. This verification module, based on a preset deformation-vibration correlation model, verifies the physical and logical consistency between the data, eliminates abnormal data segments caused by sensor failure or environmental interference, and improves the reliability of the monitoring data.

[0083] The above-described arrangement, through spatial symmetry, directional calibration optimization, and multi-source data verification, constructs a high-precision collaborative monitoring system for furnace deformation and vibration. The combined configuration of a normal displacement sensor and a three-dimensional dilatometer enables multi-dimensional analysis of the deformation vector; the layered frequency domain feature extraction of the vibration monitoring unit enhances the identification capability of combustion dynamic features.

[0084] Specifically, the combustion fine adjustment method based on full furnace expansion monitoring, wherein the normal deformation, three-dimensional expansion displacement data and vibration spectrum data are respectively processed by wavelet noise reduction filtering and temperature compensation to generate a standardized monitoring dataset including displacement features and vibration features, includes: using a wavelet transform algorithm to filter out high-frequency noise from the original vibration spectrum data collected by the vibration monitoring unit;

[0085] The temperature compensation process includes: linearly correcting the expansion displacement measurement value obtained by the expansion sensor array based on the furnace ambient temperature data to eliminate the deviation of the thermal expansion coefficient.

[0086] When performing wavelet denoising on the raw vibration spectrum data collected by the vibration monitoring unit, a suitable wavelet basis function is first selected based on the frequency domain characteristics of the combustion system vibration signal. The signal is then decomposed into high-frequency detail components and low-frequency approximate components through multi-scale decomposition. For the noise band in the high-frequency components generated by mechanical transmission or electromagnetic interference, an adaptive threshold algorithm is used to nonlinearly shrink the detail coefficients, preserving the effective vibration characteristics related to combustion pulsation. The processed wavelet coefficients at each level are then reconstructed using a reconstruction algorithm to generate denoised vibration spectrum data, suppressing the interference of high-frequency noise on combustion state analysis.

[0087] The selection of wavelet basis functions is based on the typical frequency distribution characteristics of the burner vibration signal, prioritizing basis functions with excellent time-frequency localization characteristics to match the transient impact characteristics of the combustion vibration signal. The number of decomposition levels is set based on the resonant frequency range of the combustion chamber structure, and the optimal decomposition level that can effectively separate combustion-related vibration modes is determined through preliminary experiments. Modulus maxima detection is introduced during thresholding to distinguish noise components from energy abrupt changes in actual vibration events, avoiding excessive attenuation of effective signal components.

[0088] For the measurement data from the expansion sensor array, temperature compensation processing requires establishing a correlation model between the furnace ambient temperature and the sensor's thermal expansion coefficient. Temperature sensing nodes are synchronously deployed at the sensor installation locations to collect real-time data on the furnace wall surface temperature distribution. Based on the material's thermal expansion characteristic curve, a linear regression model of the expansion displacement measurement value and temperature change is constructed, and the temperature compensation coefficient matrix is ​​fitted using the least squares method. In the data preprocessing stage, the real-time temperature data collected by the temperature sensing nodes is input into the compensation model to dynamically correct the original displacement measurement values ​​of the expansion sensors, eliminating measurement reference drift caused by furnace temperature fluctuations.

[0089] In the temperature compensation model, differentiated compensation parameters are set for different furnace wall structural zones based on their materials. For the water-cooled wall tube region, the thermal expansion influence weights of metallic components and non-metallic sealing materials are calculated separately, taking into account the difference in thermal conductivity between the tubes and fins. The compensation coefficient matrix is ​​dynamically updated according to the furnace operating conditions. When the temperature gradient change inside the furnace exceeds a set threshold, a recalibration process for the compensation parameters is triggered to maintain the long-term stability of displacement measurements. After the compensated displacement data and vibration spectrum data are timestamped, they are input into the standardization processing module to generate a monitoring dataset with a unified time reference and physical dimensions.

[0090] The above processing flow optimizes the accuracy of vibration feature extraction through signal decomposition and reconstruction techniques, and enhances the environmental adaptability of expansion measurements by combining a temperature compensation mechanism. Wavelet denoising preserves the time-frequency characteristics of combustion-related vibration modes, providing high-quality input for subsequent spectral analysis; the temperature compensation algorithm eliminates the systematic bias of thermal expansion effects on deformation measurements, enhancing the physical consistency of displacement data. The synergistic application of these two preprocessing techniques lays the data foundation for flame center offset modeling and anomalous region identification.

[0091] Specifically, the combustion fine adjustment method based on full furnace expansion monitoring involves calculating the expansion deviation rate of symmetrical monitoring points based on the three-dimensional expansion displacement of the four furnace walls at the same height level in the standardized monitoring dataset, constructing a flame center offset vector model using a random forest algorithm based on the expansion deviation rate, outputting a heat load distribution cloud map reflecting the spatial distribution of heat load, and calculating the expansion deviation rate of symmetrical monitoring points at the same height level based on the expansion displacement of the four furnace walls in the standardized monitoring dataset.

[0092] The mapping relationship between the expansion deviation rate and the flame center coordinates is established by using the random forest algorithm, and a heat load spatial distribution matrix with weight coefficients is generated.

[0093] Based on a standardized monitoring dataset, three-dimensional expansion displacement vector decomposition is performed on symmetrical monitoring points on the four furnace walls at the same height level, extracting the axial, radial, and tangential displacement components of each monitoring point in the furnace coordinate system. By calculating the difference ratio of the three-dimensional displacement vector magnitudes of the symmetrical monitoring point pairs and combining it with the cosine value of the displacement direction angle, a set of expansion deviation rate parameters characterizing the asymmetric deformation of the four furnace walls is generated. The selection of the symmetrical monitoring point pairs is based on the geometric symmetry of the furnace, with each pair containing two sets of sensor nodes arranged at 180-degree intervals along the circumference of the furnace, covering four sets of orthogonal symmetry axes in the burner area.

[0094] The expansion deviation rate was calculated using a normalization process, converting the differences in the three-dimensional displacement components into dimensionless proportionality coefficients to eliminate the dimensional influence caused by differences in the absolute positions of the monitoring points. For each height layer, four pairs of symmetrical monitoring points were used to generate three types of characteristic parameters: axial expansion difference rate, radial offset angle, and tangential displacement gradient, forming a set of characteristic vectors characterizing the asymmetry of the heat load distribution at that layer.

[0095] The random forest algorithm's model training is based on the correspondence between the expansion deviation rate feature vector and the measured coordinates of the flame center in historical operating datasets. Input features include the axial expansion difference rate, radial offset angle, and tangential displacement gradient at each height level. The output label is the three-dimensional coordinates of the flame center calibrated by an infrared thermal imager under the corresponding operating condition. During the feature selection stage, the Gini coefficient is used to evaluate the contribution of each feature to the flame center location prediction, and redundant features are removed to optimize the model's generalization ability. The trained random forest model can output the predicted coordinates of the flame center within the furnace space and a confidence index based on the real-time expansion deviation rate feature vector.

[0096] The heat load spatial distribution matrix is ​​constructed based on the flame center coordinate prediction results, and the heat load density distribution of the furnace cross-section is generated using a radial basis function interpolation algorithm. The weight coefficients of the matrix are assigned according to the furnace heat transfer characteristics, with higher weights assigned to grid cells near the flame center and weights in the edge regions decreasing exponentially with distance. The matrix update frequency is synchronized with the sensor data acquisition cycle, enabling a dynamic and visual representation of the heat load distribution. The dynamic adjustment mechanism of the weight coefficients incorporates the number of burner operating layers and the secondary damper opening as correction factors to enhance the model's adaptability to changes in operating conditions.

[0097] The model validation phase assesses the accuracy of the flame center offset vector model by comparing the consistency between the predicted heat load distribution and the boiler thermal calculation results. When the deviation between real-time monitoring data and model predictions exceeds a preset threshold, an online optimization process for model parameters is triggered. An incremental learning algorithm is used to update the split nodes of the random forest decision tree to maintain the model's prediction accuracy under varying operating conditions. This process establishes a quantitative relationship between furnace thermal state and structural deformation by integrating mechanical deformation monitoring data with machine learning algorithms.

[0098] Specifically, the refined combustion adjustment method based on full furnace expansion monitoring includes extracting the coordinates of monitoring points whose normal deformation exceeds the material's elastic limit from the standardized monitoring dataset, associating them with the energy proportion features of the low-frequency band in the vibration spectrum data at the corresponding locations, and identifying local heat flux density anomaly regions using a support vector machine classifier.

[0099] The coordinates of monitoring points whose normal deformation exceeds the elastic limit of the water-cooled wall tube are extracted from the standardized monitoring dataset. The elastic limit is dynamically set based on the yield strength threshold of the water-cooled wall tube and in combination with the real-time monitored furnace temperature.

[0100] By combining the temporal energy distribution characteristics in the vibration spectrum data at the corresponding location, a support vector machine classifier is used to determine the risk level of combustion instability.

[0101] A set of anomaly region coordinates, including priority labels, is generated and used as input parameters for the reinforcement learning model.

[0102] In the process of processing standardized monitoring datasets, when dynamically setting the elastic limit threshold of water-cooled wall pipes, a temperature compensation factor matrix is ​​established based on the yield strength characteristic curve of the pipes and the real-time collected furnace temperature field distribution data. The elastic modulus attenuation coefficient of the pipes under different temperature gradients is calculated using a material thermodynamic model, and the elastic limit judgment threshold at each monitoring point is dynamically adjusted to make the deformation exceeding the limit judgment more consistent with actual working conditions. For areas with uneven temperature distribution, a weighted average algorithm of nearby monitoring point temperatures is used to correct local thresholds, avoiding threshold setting deviations caused by sparse temperature measurement points.

[0103] When extracting the coordinates of monitoring points with excessive normal deformation, a multi-dimensional sliding window scan is performed on the standardized dataset to identify monitoring points where the deformation exceeds the dynamic threshold for three consecutive sampling periods. For the identified anomalies, the temporal energy distribution data of the vibration spectrum at the corresponding location is retrieved simultaneously, and the energy entropy value and low-frequency energy ratio characteristics of the vibration signal within a preset time window are extracted. The length of the time window is set to match the combustion adjustment cycle, so that the vibration characteristics can reflect the cumulative effect of changes in the combustion state.

[0104] The training sample set for the support vector machine classifier includes vibration spectrum feature vectors corresponding to deformation exceeding events during historical operation, as well as combustion instability state labels verified by thermodynamic calculations. During feature vector construction, the deformation exceeding amplitude, vibration energy entropy, and low-frequency energy proportion are normalized to form a multi-dimensional input space. The choice of the classifier kernel function is based on the linear separability test results of the feature vectors, preferentially using radial basis functions to achieve high-dimensional space mapping. A cross-validation mechanism is introduced during training to optimize the classification boundary and improve the ability to distinguish between minor combustion disturbances and severe instability states.

[0105] The classifier output includes a risk level label and a confidence score. The risk level is determined based on the rate of plastic deformation of the pipe that may be caused by combustion instability. High-risk anomalies with high confidence are marked as first-priority in the coordinate set; anomalies with medium to low risk but with consecutive multi-cycle trigger records are marked as second-priority. The priority marking rule base integrates expert experience and knowledge; when three or more adjacent anomalies appear at the same height layer, the priority level of that area is automatically increased. The marked coordinate set is bound to a timestamp and risk level code to form a structured input parameter format.

[0106] The aforementioned identification process enhances environmental adaptability through dynamic threshold adjustment, improves the reliability of anomaly detection through multi-source data correlation analysis, and optimizes resource allocation through a priority labeling mechanism. Dynamic calculation of the elastic limit addresses temperature field interference; vibration time-domain energy feature extraction captures combustion dynamics; and support vector machine classification enables multi-dimensional feature fusion decision-making.

[0107] Specifically, the workflow of the reinforcement learning model in the combustion fine-tuning method based on full furnace expansion monitoring includes:

[0108] The generated spatial distribution matrix of heat load is spatially encoded with the set of coordinates of abnormal areas;

[0109] The data include NOx concentration, oxygen content, furnace outlet temperature, desuperheating water volume, and steam temperature from the boiler operating parameters.

[0110] The combination of adjustment parameters under the multi-objective constraints of heat load balancing, NOx emission minimization, and steam temperature stability is solved by a near-end strategy optimization algorithm.

[0111] In the spatial encoding of the heat load spatial distribution matrix, a three-dimensional furnace mesh generation technique is employed to map the weight coefficients in the matrix to corresponding physical coordinate grid cells. The heat load value of each grid cell is converted into state vector elements through normalization, and the coordinate set of abnormal regions is marked with an isolated thermal encoding method to generate a two-dimensional encoding matrix containing information on heat load intensity and abnormal location. A spatiotemporal dimension expansion mechanism is introduced during the encoding process, superimposing the heat load change gradients of three adjacent sampling periods as time series features into the spatial encoding matrix to enhance the model's ability to understand the dynamic combustion process.

[0112] The fusion processing of boiler operating parameters employs a multi-source data standardization method. Sliding window mean and variance normalization are applied to NOx concentration, oxygen content, furnace outlet temperature, desuperheating water flow, and steam temperature to eliminate the influence of dimensional differences on the model input. The standardized parameter data is synchronized with the spatial encoding matrix via a timestamp alignment module to construct a three-channel input tensor containing heat load distribution, abnormal region markers, and operating parameters. Before being input into the reinforcement learning model, this tensor undergoes a feature importance evaluation module to filter key feature dimensions, reducing data redundancy from interfering with strategy optimization.

[0113] In the implementation of the near-end policy optimization algorithm, a composite reward function is defined, including heat load balance, NOx emission intensity, furnace outlet temperature gradient, desuperheating water flow gradient, and steam temperature deviation. Heat load balance is quantified by calculating the load difference entropy between adjacent grids in the spatial encoding matrix; NOx emission intensity is evaluated using the cumulative emission gradient within a sliding window; and steam temperature deviation is calculated based on the standard deviation of the set value. The policy network, through a gradient update mechanism under trust domain constraints, explores the action space of burner output adjustment and damper opening combinations in each iteration, seeking a Pareto optimal solution set under multi-objective constraints.

[0114] During strategy optimization, an action masking mechanism is introduced to transform the physical constraints of the burner actuator into feasible domain boundaries of the action space. For damper opening adjustment, single-step adjustment range limits are set based on the mechanical response characteristics of the actuator; burner output adjustment is correlated with the flow control margin of the fuel supply system, generating discrete action options that conform to actual control capabilities. The optimized adjustment parameter combinations undergo a feasibility verification module to eliminate aggressive strategies that may cause transient shocks to the equipment, generating a safe and controllable adjustment command sequence.

[0115] The above workflow achieves digital reconstruction of combustion state through spatial coding, establishes global optimization objectives through multi-source data fusion, and balances multi-dimensional control requirements through near-end strategy optimization. The coding mechanism solves the problem of mathematical representation of spatial distribution of heat load and abnormal regions; data standardization and feature screening improve the quality of model input; and the design of composite reward functions and action constraints ensures the engineering applicability of the control strategy. The technical features of each step serve the in-depth application of mechanical deformation monitoring data and the generation of combustion control decisions.

[0116] Specifically, the refined combustion adjustment method based on full furnace expansion monitoring includes the real-time acquisition of water-cooled wall expansion rate change data within a preset time period after adjustment, optimization of the weight parameters of the reinforcement learning model using gradient descent based on the expansion rate change data, and updating the control strategy in the combustion adjustment rule base.

[0117] Collect data on the change in the water-cooled wall expansion rate within a preset time range after the burner actuator completes its adjustment action;

[0118] Calculate the expansion suppression efficiency coefficient of the current adjustment strategy based on the expansion rate change data;

[0119] Based on the inflation suppression efficiency coefficient, the weight parameters of the reinforcement learning model are optimized and the reward function value is updated using the gradient descent method.

[0120] After the burner actuator completes its adjustment, a dynamic response time window is set to collect data on the expansion rate changes of the water-cooled wall. The start time of the time window is synchronized with the time the adjustment command is issued, and the window length is dynamically adjusted according to the current boiler load rate. Under high load conditions, the acquisition cycle is shortened to match the rapid thermal response characteristics. During data acquisition, the expansion rate at the same monitoring point is cross-validated using redundant sensor nodes to eliminate abnormal data segments caused by local sensor failures, thereby improving the reliability of the dataset.

[0121] The expansion suppression efficiency coefficient is calculated using a multi-dimensional evaluation method, decomposing the expansion rate change curve within the time window into a transient response stage and a steady-state convergence stage. In the transient stage, the effective time of the adjustment strategy's suppression of mechanical deformation is identified by calculating the extreme points of the second derivative of the rate change curve. In the steady-state stage, the percentage of the area experiencing a rate decrease is obtained through integral calculations, and combined with the heat transfer safety weight allocation coefficients for different regions of the water-cooled wall, a comprehensive efficiency coefficient reflecting the adjustment effect is generated. The weight allocation coefficients are dynamically set based on the structural stress distribution diagram of the water-cooled wall tube panel, with higher weights assigned to the suppression effect in high heat load regions.

[0122] In the gradient descent optimization process, the expansion suppression efficiency coefficient is used as a negative penalty term in the policy network loss function, forming a multi-objective optimization problem with the heat load balance and NOx emission indicators in the original reward function. An adaptive learning rate adjustment mechanism is employed during optimization iterations, dynamically scaling the parameter update step size based on historical gradient magnitudes to avoid getting trapped in local optima. After updating the model weights, the coefficients of each objective term in the reward function are normalized and redistributed to increase the priority of expansion suppression in policy decision-making, while maintaining a balance between pollutant emission control, combustion flow field, and steam temperature stability.

[0123] The updated model parameters are verified through offline simulation and then imported into the online rule base to replace the original policy entries. During the verification process, a virtual combustion scenario is constructed, and typical disturbance patterns from historical operating data are injected to test the generalization ability of the new strategy under various operating conditions. For policy entries that pass verification, their corresponding efficiency coefficient improvement and applicable operating condition range are recorded as the priority ranking basis for subsequent policy calls. The above process forms a closed-loop control architecture of "execution-monitoring-evaluation-optimization".

[0124] Specifically, the refined combustion adjustment method based on full furnace expansion monitoring also includes:

[0125] The expansion displacement pattern, vibration spectrum characteristics, and combustion parameters from historical operating data are combined and input into a convolutional neural network.

[0126] Spatiotemporal correlation feature vectors are extracted and dimensionality reduced by principal component analysis to generate a combustion health index.

[0127] The combustion state is classified into three categories based on the combustion health index: steady state, wave dynamics, and unstable state, and the classification results are fed back to the reinforcement learning model.

[0128] In the preprocessing stage of historical operating data, expansion displacement patterns, vibration spectrum characteristics, and combustion parameters are combined to construct a three-dimensional spatiotemporal data cube. The time dimension of the cube is sliced ​​according to the combustion adjustment cycle, the spatial dimension corresponds to the furnace monitoring grid coordinates, and the channel dimension includes parameters such as displacement vector, vibration energy spectrum, and oxygen content. Before the data cube is input into the convolutional neural network, sliding window normalization is used to eliminate differences in the measurement range of different sensors, and zero-filling technology is used to maintain the continuity of the spatial dimension.

[0129] The convolutional neural network architecture incorporates parallel extraction paths for spatiotemporal features. The dilation displacement pattern captures spatial distribution characteristics via a two-dimensional convolutional layer, while the vibration spectrum time series extracts dynamic evolution patterns through a one-dimensional temporal convolutional layer. After deep fusion of these two types of features, a three-dimensional max-pooling layer is used to compress the data dimension, generating feature vectors that incorporate the spatiotemporal correlation characteristics of the furnace thermal state. The step size of the pooling layer is matched to the combustion adjustment cycle, enabling the feature vectors to characterize the key state transition nodes of the complete combustion process.

[0130] In the principal component analysis (PCA) dimensionality reduction process, singular value decomposition is performed on the covariance matrix of the eigenvectors, and eigenvectors with a cumulative contribution rate exceeding a set threshold are selected to construct the projection space. The dimensionality-reduced low-dimensional vectors are then fused using linear weighting to generate a combustion health index. The weight allocation is dynamically adjusted based on the correlation analysis results between the eigenvectors and historical failure events. The scaling range of the health index is set according to the boiler design parameters, with the lower limit corresponding to the ideal stable combustion state and the upper limit related to the critical condition of pipe plastic deformation.

[0131] Combustion state classification employs a Gaussian mixture model clustering algorithm, dividing the time-series changes in the combustion health index into three categories: steady state, wave dynamics, and unstable state. The classification boundary is iteratively optimized using an expectation-maximization algorithm, and cluster centers are manually calibrated based on expert experience to eliminate misclassifications caused by uneven data distribution. The classification results are fed back to the reinforcement learning model in the form of state codes, serving as additional conditional variables in the policy network input layer, guiding the model to prioritize historically validated adjustment strategies under specific combustion states.

[0132] The above process enhances state representation capabilities through spatiotemporal data fusion, improves feature engineering efficiency through principal component analysis, and achieves hierarchical description of combustion states through clustering and classification. Convolutional neural network feature extraction solves the problem of fusion and analysis of multi-source heterogeneous data; health index constructs a quantitative analysis of the overall stability of the combustion system; and state classification feedback mechanism enhances the environmental adaptability of the reinforcement learning model.

[0133] Specifically, the refined combustion adjustment method based on full furnace expansion monitoring also includes:

[0134] Based on coal quality change data, a transfer learning algorithm is invoked to adapt the parameters of the reinforcement learning model.

[0135] The kernel function parameters of the support vector machine classifier are dynamically adjusted based on the classification results of the combustion health index.

[0136] The combustion health index determination threshold of the convolutional neural network is updated based on the range of changes in coal calorific value.

[0137] In the processing of coal quality change data, a mapping relationship library between coal quality feature vectors and combustion monitoring datasets is established. The feature vectors contain coal quality industrial analysis parameters and elemental analysis parameters, and valid sample data are screened using an outlier detection algorithm. When a deviation of the current coal quality feature vector from the historical data distribution is detected, a transfer learning process is triggered. The parameters of the policy network layer in the reinforcement learning model that are strongly correlated with fuel characteristics are frozen, while the general parameters of the thermodynamic equilibrium constraint layer are retained. Incremental training is then performed using limited sample data under new coal quality conditions to achieve rapid adaptation of model parameters.

[0138] During transfer learning, feature alignment techniques are employed, and the maximum mean difference algorithm is used to reduce the distributional differences between new and old coal quality data in the combustion feature space. The trained adaptation model retains its original multi-objective optimization capabilities while enhancing its sensitivity to the current coal quality combustion characteristics. In the model validation phase, gradually changing coal quality test sequences are injected to observe the smooth transition characteristics of the policy network's output adjustment parameters, preventing sudden parameter changes from causing actuator oscillations.

[0139] A dynamic adjustment mechanism for the kernel function parameters of the support vector machine classifier is implemented, constructing a feedback loop based on the classification results of the combustion health index. When the health index remains stable, a larger kernel width parameter is used to enhance the fault tolerance of the classification boundary; when the index enters a volatile or unstable state, the kernel width is reduced and the penalty factor weight is increased to improve the accuracy of identifying abnormal features. The magnitude of parameter adjustment is positively correlated with the degree to which the health index deviates from the baseline value. An exponential moving average algorithm is used to smooth the adjustment process and avoid frequent parameter jumps.

[0140] During kernel function parameter adjustment, the feature space mapping relationship is simultaneously optimized. For feature dimensions with low contribution to the combustion health index, their weight ratio in kernel function calculation is gradually reduced to strengthen the influence of key features on classification decisions. This optimization process works in conjunction with the feature alignment module of transfer learning to form a classifier architecture with adaptive combustion state sensitivity, improving the stability of anomaly detection under different coal quality conditions.

[0141] The convolutional neural network-based threshold update strategy for combustion health index determination divides the dynamic adjustment stages according to the range of coal calorific value variation. When the coal calorific value is within the design coal type range, a baseline threshold is used for state classification; when the calorific value deviates beyond a preset window, a compensation coefficient is calculated based on the calorific value-expansion rate correlation curve, and the determination threshold is scaled proportionally. A hysteresis mechanism is introduced during the threshold update process to prevent threshold oscillations caused by short-term fluctuations in coal quality, while a maximum adjustment range limit is set to maintain the physical consistency of the health index evaluation system.

[0142] The above technical process constructs a combustion monitoring and control system that adapts to fuel characteristics through a collaborative mechanism involving coal quality characteristics-driven model parameter adaptation, health status feedback to optimize classification accuracy, and calorific value-related threshold updates. Transfer learning maintains the model's generalization ability to new coal types, dynamic adjustment of the kernel function enhances anomaly detection sensitivity, and a threshold compensation mechanism ensures the reliability of state classification.

[0143] Specifically, the refined combustion adjustment method based on full furnace expansion monitoring further includes: after each combustion adjustment command is executed, collecting expansion displacement correlation data of two adjacent monitoring points arranged at preset intervals along the circumference of the furnace in the same height layer, and four monitoring points in the upper and lower adjacent layers.

[0144] Based on the aforementioned correlation data, a transfer function relating vibration spectrum characteristics to changes in heat flux density is constructed.

[0145] The heat load weighting coefficient of the flame center offset vector model is corrected by the transfer function.

[0146] After the combustion adjustment command is executed, when collecting expansion displacement correlation data for a preset combination of monitoring points, the spatial layout of the monitoring points adopts a cross-layer coupling design. Two adjacent monitoring points arranged at intervals along the circumference of the furnace in the same height layer form a horizontal correlation pair, and four monitoring points in adjacent upper and lower layers form a pyramid-shaped spatial correlation group. The data acquisition module synchronously records the time-domain waveform of the three-dimensional expansion displacement of each monitoring point, and calculates the phase difference and amplitude ratio of the displacement waveform through a cross-correlation algorithm to generate a set of correlation parameters characterizing the thermal expansion propagation characteristics.

[0147] The processing of correlation data employs a system identification algorithm, which performs time-delay matching between the dominant frequency components in the vibration spectrum characteristics and the rate of change of heat flux density. When constructing the transfer function, the vibration acceleration signal is used as the input variable, and the gradient of heat flux density change is used as the output response. The coefficient matrix of the transfer function is solved using the recursive least squares method. The transfer function includes two dimensions: amplitude-frequency characteristics and phase-frequency characteristics, reflecting the dynamic coupling relationship between mechanical vibration and thermal load fluctuations under specific combustion conditions.

[0148] The transfer function's correction to the flame center offset vector model is achieved through dynamic compensation of the heat load weighting coefficients. Within the update cycle of the heat load spatial distribution matrix, the resonant peak frequency components in the transfer function's amplitude-frequency characteristics are extracted to identify heat flux density-sensitive frequency bands. Based on the energy distribution intensity of these sensitive frequency bands, gain compensation is applied to the heat load weighting coefficients of the corresponding spatial grids to enhance the correction effect of vibration characteristics on the heat load distribution. After normalization, the corrected weighting coefficients are redistributed to the feature importance evaluation module of the random forest algorithm, forming a two-way coupling mechanism for the model parameters.

[0149] The verification of the correction effect employs a virtual combustion disturbance test method. A standard heat flux step signal is injected into the simulation environment, and the prediction deviation of the flame center coordinates before and after the transfer function correction is compared. When the improvement in prediction accuracy of the corrected model reaches a preset threshold, the current transfer function parameters are fixed to the model library as benchmark correction parameters for similar combustion conditions. The closed-loop correction process continuously optimizes the analytical accuracy of heat load distribution through online identification of vibration-heat flux correlation characteristics.

[0150] The following is an explanation of the main technical feature terms in the technical solution of this invention:

[0151] Expansion sensor array: A monitoring device composed of normal displacement sensors and a three-dimensional dilatometer, arranged at multiple height levels on the four walls of the furnace according to a preset geometric pattern. The normal displacement sensors are installed along the normal direction of the furnace walls to measure deformation perpendicular to the furnace wall plane; the three-dimensional dilatometer synchronously acquires axial, radial, and tangential displacement components through a multi-axis sensing structure, forming a three-dimensional deformation monitoring capability. This array achieves high-density capture of the furnace's full circumferential deformation characteristics through a spatially symmetrical layout and a redundancy verification mechanism.

[0152] Vibration monitoring unit: A monitoring layer consisting of vibration acceleration sensors and a spectrum analysis module, arranged in layers from the main burner to the SOFA wind area. The vibration acceleration sensors are installed at a specific tilt angle to optimize signal sensitivity in the direction of combustion disturbance; the spectrum analysis module performs time-frequency domain windowing processing on the raw vibration signal to extract characteristic parameters such as energy spectrum and kurtosis, reflecting the coupling effect between combustion dynamics and mechanical vibration.

[0153] Wavelet noise reduction: A signal preprocessing method based on wavelet transform separates vibration spectrum data into high-frequency noise components and low-frequency effective signals through multi-scale decomposition. An adaptive threshold algorithm is used to nonlinearly shrink the high-frequency coefficients, preserve the time-frequency characteristics of combustion-related vibration modes, and suppress the interference of electromagnetic interference and mechanical noise on subsequent analysis.

[0154] Temperature compensation processing: The displacement measurements of the expansion sensors are dynamically corrected based on real-time data of the furnace wall surface temperature. A linear compensation model is established based on the material's thermal expansion coefficient curve to eliminate measurement reference drift caused by ambient temperature fluctuations, ensuring physical consistency between the displacement data and the actual deformation state.

[0155] Flame center offset vector model: This machine learning model, constructed using the random forest algorithm, takes as input the expansion deviation rate feature vector of symmetrical monitoring points at the same height layer, and outputs the predicted three-dimensional coordinates of the flame center and the spatial distribution matrix of the heat load. The model establishes a mapping relationship between the expansion deviation rate and the heat load distribution through training with historical data, achieving sub-meter-level heat flux density analysis.

[0156] Support Vector Machine Classifier: A supervised learning model based on kernel function mapping. It takes the vibration spectrum low-frequency energy features of the deformation exceeding the limit monitoring point as input, constructs the optimal classification hyperplane through the high-dimensional feature space, determines the risk level of combustion instability in the local area, and generates a set of anomaly coordinates with priority labels.

[0157] Reinforcement learning model: A decision-making model employing a proximal policy optimization algorithm integrates heat load distribution, coordinates of abnormal regions, and boiler operating parameters to generate adjustment commands for burner output and damper opening under multi-objective constraints. The model quantifies the effects of heat load balance, emission control, and deformation suppression through a reward function, driving the policy network to explore the optimal action space.

[0158] Transfer function correction: A dynamic correction mechanism based on the correlation data between vibration spectrum and heat flux density. A vibration-heat flux transfer function is established through a system identification algorithm, and energy characteristics of sensitive frequency bands are extracted to compensate for the gain of the heat load weighting coefficient, thereby optimizing the dynamic response accuracy of the flame center offset model.

[0159] An array of expansion sensors is installed on each of the four furnace walls in the main burner area of ​​the boiler furnace. Each sensor node group includes a normal displacement sensor and a three-dimensional dilatometer, symmetrically distributed along the circumference of the furnace to form a deformation monitoring network covering the entire height. The normal displacement sensor is coupled to the furnace wall surface through a rigid bracket to collect deformation perpendicular to the furnace wall plane in real time; the three-dimensional dilatometer adopts a multi-axis sensing structure to simultaneously record axial, radial, and tangential displacement components. Vibration monitoring units are arranged in layers from the main burner to the SOFA wind area. Vibration acceleration sensors are installed at a preset angle to capture the characteristic spectrum of the combustion disturbance direction. After orientation calibration and coordinate system mapping, the sensor nodes are eliminated by a synchronous sampling module to generate a spatiotemporally consistent deformation and vibration dataset.

[0160] In the raw data preprocessing stage, the vibration spectrum data is decomposed into multiple scales using wavelet basis functions. An adaptive threshold algorithm is used to filter out high-frequency mechanical noise while retaining low-frequency pulsation characteristics related to combustion. Expansion sensor data is combined with real-time monitoring values ​​of the furnace wall surface temperature. Displacement measurements are dynamically corrected based on the material's thermal expansion coefficient to eliminate reference drift caused by temperature gradients. The preprocessed standardized dataset is then input into a random forest algorithm to calculate the three-dimensional expansion deviation rate of symmetrical monitoring points at the same height layer, establishing a mapping model between the expansion deviation rate and the spatial position of the flame center. This model is trained using historical operating data and outputs a heat load distribution cloud map with weighted coefficients, analyzing the sub-meter-level heat flux density distribution of the furnace cross-section.

[0161] When identifying local anomaly areas, the elastic limit threshold of the water-cooled wall pipe is dynamically calculated, and the deformation exceeding the limit judgment standard is corrected in combination with the temperature field distribution. For monitoring points that continuously exceed the limit, the low-frequency energy entropy value features of the vibration spectrum at the corresponding location are extracted, and the combustion instability risk level is evaluated by a support vector machine classifier to generate a set of anomaly coordinates with priority labels. The reinforcement learning model integrates operating parameters such as heat load cloud map, anomaly area coordinates, and NOx concentration, and uses a near-end strategy optimization algorithm to generate adjustment commands for burner output and damper opening under multi-objective constraints. After the commands are executed, the water-cooled wall expansion rate change data is collected, the model weights are optimized based on the gradient descent method, and the control strategy library is updated to form a closed-loop feedback mechanism. This implementation method achieves coordinated optimization control of combustion state and mechanical deformation through dynamic correlation analysis of deformation and heat load and real-time strategy iteration.

[0162] Flame center offset vector model: A machine learning model based on the random forest algorithm. It takes as input the 3D expansion displacement deviation rate feature vector of symmetrical monitoring points at the same height layer, and establishes a nonlinear mapping relationship between the expansion deviation rate and the 3D coordinates of the flame center through an ensemble learning mechanism of multiple decision trees. The model evaluates feature importance using the Gini coefficient, filters key deformation features sensitive to flame offset, and outputs a heat load distribution cloud map with spatial weight coefficients, achieving sub-meter spatial resolution of the furnace cross-sectional heat flux density.

[0163] Support Vector Machine (SVM) Classifier: A supervised learning model employing a radial basis function kernel is used. The input parameters are the low-frequency energy entropy and energy proportion features of the vibration spectrum at the normal deformation exceeding the limit monitoring point. An optimal classification hyperplane is constructed in the high-dimensional feature space, dividing the local region into stable, transitional, and unstable risk levels. The classification boundary is optimized through cross-validation, and a set of abnormal region coordinates is generated by combining a priority labeling rule base, providing spatial constraints on combustion instability risk for the reinforcement learning model.

[0164] Reinforcement learning model: A multi-objective decision-making model based on the Proximal Policy Optimization (PPO) algorithm. Inputting heat load distribution cloud map, coordinates of abnormal areas, and boiler operating parameters, the model explores the action space of burner output and damper opening through a policy network. The model defines a composite reward function to quantify heat load balance (entropy difference between adjacent grids), NOx emission gradient, furnace outlet temperature gradient, desuperheating water gradient, and steam temperature deviation. Under trust domain constraints, it generates a sequence of adjustment instructions that meet equipment safety and environmental protection requirements, achieving dynamic optimization of combustion parameters.

[0165] Transfer function correction mechanism: Based on the vibration-heat flux dynamic coupling model constructed using system identification technology, the expansion displacement correlation data and vibration spectrum characteristics of a specific monitoring point group are input. The transfer function coefficient matrix of vibration signal and heat flux density change is solved using the recursive least squares method. By extracting the resonant peak frequency component in the amplitude-frequency characteristics of the transfer function, frequency domain gain compensation is performed on the heat load weight coefficient of the flame center offset model to eliminate analytical errors in heat load distribution caused by combustion disturbance.

[0166] Convolutional Neural Network Model: A spatiotemporal feature extraction model for generating the combustion health index. It takes as input a spatiotemporal data cube containing historical expansion displacement patterns, vibration spectra, and combustion parameters. A two-dimensional convolutional layer captures spatial deformation distribution features, while a one-dimensional temporal convolutional layer extracts the dynamic evolution of vibration. The fused high-order features are then dimensionality-reduced using principal component analysis to generate a scalarized health index characterizing the overall stability of the combustion system, providing an input benchmark for state classification.

[0167] Transfer Learning Adaptation Module: This module optimizes model parameters for changes in coal quality. It reduces the distribution differences between old and new coal quality data in the combustion feature space through feature alignment algorithms, freezes the general parameter layers related to thermodynamic equilibrium in the reinforcement learning model, and only performs incremental training on policy network layers sensitive to fuel characteristics. This module retains the generalization ability of historical policies while adapting to the combustion kinetics of the current coal quality, avoiding inaccurate control commands caused by fuel changes.

[0168] Combustion Health Index Generation Model: Based on a Gaussian mixture model clustering classification model, the model takes a dimensionality-reduced feature vector extracted by a convolutional neural network as input and iteratively optimizes the cluster centers using an expectation-maximization algorithm, classifying the combustion health index into three categories: steady state, wave dynamics, and unstable state. The classification results are fed back to the reinforcement learning model in the form of state codes, constraining the action exploration range of the policy network under specific combustion conditions and improving the condition adaptability of control commands.

[0169] This invention solves the problem of combustion adjustment lag caused by insufficient spatial resolution in existing temperature monitoring methods through the following technical solution:

[0170] Full-dimensional deformation monitoring and data fusion: Expansion sensor arrays are deployed on all four sides of each layer in the main burner area of ​​the boiler furnace, and vibration monitoring units are simultaneously set up layer by layer from the main burner to the SOFA air area. The expansion sensor arrays collect the normal deformation and three-dimensional expansion displacement data of the four furnace walls, while the vibration monitoring units acquire dynamic vibration spectrum data of combustion, constructing a collaborative monitoring network for mechanical deformation and vibration covering the entire furnace area. High-frequency noise interference is filtered out by wavelet denoising, and linear compensation is performed on the expansion measurements in combination with real-time furnace temperature data to generate a standardized displacement and vibration feature dataset, breaking through the limitations of single-dimensional temperature monitoring and realizing multi-dimensional spatial analysis of heat load distribution.

[0171] Dynamic heat load modeling and anomaly identification: Based on a standardized dataset, the expansion deviation rate of symmetrical monitoring points at the same height layer is calculated. A random forest algorithm is used to establish a mapping relationship between the expansion deviation rate and the flame center coordinates, generating a heat load spatial distribution matrix with sub-meter resolution. Simultaneously, the coordinates of monitoring points with excessive normal deformation are extracted, and low-frequency energy features of the vibration spectrum are correlated. A support vector machine classifier is used to identify local heat flux density anomaly regions. This method, through a deformation-heat load correlation model, achieves second-level dynamic tracking of flame center offset and precise location of anomaly regions, eliminating analytical errors caused by heat transfer interference in temperature monitoring.

[0172] Closed-loop combustion control and strategy optimization: The heat load distribution map, coordinates of abnormal areas, and boiler operating parameters are input into a reinforcement learning model. A near-end strategy optimization algorithm generates multi-objective optimization adjustment commands for burner output and damper opening. After adjustment, real-time data on water-cooled wall expansion rate changes are collected. The model weights are optimized using the gradient descent method, and the control strategy library is updated, forming a closed-loop control architecture of "monitoring-decision-execution-feedback." This mechanism achieves rapid response and adaptive adjustment to heat load imbalances through dynamic coupling analysis of combustion state and mechanical deformation, effectively addressing the risk of combustion instability caused by the lag in traditional temperature monitoring.

Claims

1. A method for refined combustion adjustment based on full furnace expansion monitoring, characterized in that, include: An array of expansion sensors is arranged in each layer of the main burner area of ​​the boiler furnace to collect the normal deformation and three-dimensional expansion displacement data of the four furnace walls. Vibration monitoring units are set up in layers from the main burner area to the SOFA wind area to obtain vibration spectrum data. The normal deformation, three-dimensional expansion displacement data, and vibration spectrum data are subjected to wavelet noise reduction filtering and temperature compensation processing respectively to generate a standardized monitoring dataset including displacement features and vibration features. Based on the three-dimensional expansion displacement of the four furnace walls at the same height in the standardized monitoring dataset, the expansion deviation rate of the symmetrical monitoring points is calculated. Based on the expansion deviation rate, a flame center offset vector model is constructed using the random forest algorithm, and a heat load distribution cloud map reflecting the spatial distribution of heat load is output. The coordinates of monitoring points whose normal deformation exceeds the material's elastic limit are extracted from the standardized monitoring dataset, and the energy proportion characteristics of the low-frequency band in the vibration spectrum data of the corresponding location are associated with them. Local areas with abnormal heat flux density are identified by a support vector machine classifier. The heat load distribution cloud map output by the flame center offset vector model, the coordinate information of the abnormal area identified by the support vector machine classifier, and the oxygen content, furnace outlet temperature, desuperheating water volume, NOx concentration and steam temperature data in the boiler operating parameters are input into the reinforcement learning model to generate burner output adjustment commands and damper opening combination parameters. The distributed control system sends the burner output adjustment command and damper opening combination parameters to the burner actuator. Real-time acquisition of water-cooled wall expansion rate change data within a preset time period after adjustment; optimization of the weight parameters of the reinforcement learning model based on the expansion rate change data using gradient descent method; and updating of the control strategy in the combustion adjustment rule base. The process of collecting and adjusting the water-cooled wall expansion rate change data within a preset time period in real time, optimizing the weight parameters of the reinforcement learning model based on the expansion rate change data using the gradient descent method, and updating the control strategy in the combustion adjustment rule base includes: Collect data on the change in the water-cooled wall expansion rate within a preset time range after the burner actuator completes its adjustment action; Calculate the expansion suppression efficiency coefficient of the current adjustment strategy based on the expansion rate change data; Based on the inflation suppression efficiency coefficient, the weight parameters of the reinforcement learning model are optimized and the reward function value is updated using the gradient descent method; Also includes: The expansion displacement pattern, vibration spectrum characteristics, and combustion parameters from historical operating data are combined and input into a convolutional neural network. Spatiotemporal correlation feature vectors are extracted and dimensionality reduced by principal component analysis to generate a combustion health index. The combustion state is classified into three categories based on the combustion health index: steady state, wave dynamics, and unstable state, and the classification results are fed back to the reinforcement learning model.

2. The method for refined combustion adjustment based on full furnace expansion monitoring according to claim 1, characterized in that: The method of arranging expansion sensor arrays in each layer of the main burner area of ​​the boiler furnace to collect the normal deformation and three-dimensional expansion displacement data of the four furnace walls, and setting vibration monitoring units in layers from the main burner area to the SOFA wind area to obtain vibration spectrum data includes: Four sets of sensor nodes are set at each height level in the main burner area on the four sides of the furnace. Each set of nodes includes a normal displacement sensor and a three-dimensional dilatometer, which are used to collect the normal deformation and three-dimensional dilatation displacement data of the four furnace walls. The vibration monitoring unit is arranged in layers from the main burner area to the SOFA wind area. Each layer is equipped with a vibration acceleration sensor and a spectrum analysis module to obtain vibration spectrum data of the corresponding area.

3. The method for refined combustion adjustment based on full furnace expansion monitoring according to claim 1, characterized in that: The process of performing wavelet noise reduction filtering and temperature compensation on the normal deformation data, three-dimensional expansion displacement data, and vibration spectrum data to generate a standardized monitoring dataset including displacement and vibration features includes: The wavelet transform algorithm is used to filter out high-frequency noise from the raw vibration spectrum data collected by the vibration monitoring unit. The temperature compensation process includes: linearly correcting the expansion displacement measurement value obtained by the expansion sensor array based on the furnace ambient temperature data, in order to eliminate the deviation of the thermal expansion coefficient.

4. The method for refined combustion adjustment based on full furnace expansion monitoring according to claim 1, characterized in that: Based on the three-dimensional expansion displacement of the four furnace walls at the same height in the standardized monitoring dataset, the expansion deviation rate of the symmetrical monitoring points is calculated. Based on the expansion deviation rate, a flame center offset vector model is constructed using the random forest algorithm, and a heat load distribution cloud map reflecting the spatial distribution of heat load is output. Based on the expansion displacement of the four furnace walls in the standardized monitoring dataset, the expansion deviation rate of symmetrical monitoring points at the same height layer is calculated. The mapping relationship between the expansion deviation rate and the flame center coordinates is established by using the random forest algorithm, and a heat load spatial distribution matrix with weight coefficients is generated.

5. The method for refined combustion adjustment based on full furnace expansion monitoring according to claim 1, characterized in that: The process of extracting the coordinates of monitoring points whose normal deformation exceeds the material's elastic limit from the standardized monitoring dataset, associating them with the energy proportion characteristics of the low-frequency band in the vibration spectrum data at the corresponding locations, and identifying local heat flux density anomaly regions using a support vector machine classifier includes: The coordinates of monitoring points whose normal deformation exceeds the elastic limit of the water-cooled wall tube are extracted from the standardized monitoring dataset. The elastic limit is dynamically set based on the yield strength threshold of the water-cooled wall tube and in combination with the real-time monitored furnace temperature. By combining the temporal energy distribution characteristics in the vibration spectrum data at the corresponding location, a support vector machine classifier is used to determine the risk level of combustion instability. A set of anomaly region coordinates, including priority labels, is generated and used as input parameters for the reinforcement learning model.

6. The method for refined combustion adjustment based on full furnace expansion monitoring according to claim 5, characterized in that: The reinforcement learning model includes: The generated spatial distribution matrix of heat load is spatially encoded with the set of coordinates of abnormal areas; The data include NOx concentration, oxygen content, furnace outlet temperature, desuperheating water volume, and steam temperature from the boiler operating parameters. The combination of adjustment parameters under the multi-objective constraints of heat load balancing, NOx emission minimization, and steam temperature stability is solved by a near-end strategy optimization algorithm.

7. The method for refined combustion adjustment based on full furnace expansion monitoring according to claim 1, characterized in that, Also includes: Based on coal quality change data, a transfer learning algorithm is invoked to adapt the parameters of the reinforcement learning model. The kernel function parameters of the support vector machine classifier are dynamically adjusted based on the classification results of the combustion health index. The combustion health index determination threshold of the convolutional neural network is updated according to the range of coal calorific value variation. It also includes: after each combustion adjustment command is executed, collecting expansion displacement correlation data of two adjacent monitoring points arranged at preset intervals along the circumference of the furnace in the same height layer, and four monitoring points in the upper and lower adjacent layers; Based on the aforementioned correlation data, a transfer function relating vibration spectrum characteristics to changes in heat flux density is constructed. The heat load weighting coefficient of the flame center offset vector model is corrected by the transfer function.

8. The method for refined combustion adjustment based on full furnace expansion monitoring according to claim 1, characterized in that: The arrangement of the expansion sensor array includes a wire distance measuring hook (1), a wire distance measuring sensor (2), and a wire distance measuring fixture (3). The wire distance measuring hook (1) is fixed to the surface of the furnace outer wall insulation sheet (16) at a preset interval. The wire distance measuring sensor (2) is connected to the hook (1) through a wire to form a distance measuring grid, which is used to collect the normal deformation and three-dimensional expansion displacement data of the four furnace walls in real time. The vibration monitoring unit includes a vibration measurement module (8) and a vibration measurement probe (11). The vibration measurement probe (11) is installed on the surface of the furnace water-cooled wall (12) through the fixing bracket (13) of the whole furnace expansion monitoring device to obtain vibration spectrum data of the corresponding area. The computing control host module (7) receives real-time data from the wire distance sensor (2), vibration measurement probe (11) and wall temperature measurement probe (10) through the communication antenna (6), and generates combustion optimization instructions based on the load conditions, coal quality parameters and air volume data.

Images