A furnace temperature field virtual-real fusion reconstruction method based on digital twinning

Abstract

The present application relates to the technical field of deep learning temperature field reconstruction, and provides a furnace temperature field virtual-real fusion reconstruction method based on digital twinning. The method constructs a furnace temperature field prior benchmark through multi-working condition three-dimensional physical mechanism simulation, dynamically calibrates the virtual temperature field by introducing unit online operation data and furnace wall temperature measured values, realizes online consistency constraint of the simulation model and the actual operation state; on this basis, a deep learning temperature correction model oriented to the three-dimensional space of the furnace is constructed, and a low-rank adaptive parameter training mechanism combined with curvature perception regularization is used to realize fine correction and stable fusion of the spatial deviation of the furnace temperature field. Through the above virtual-real fusion reconstruction method, the present application realizes continuous reconstruction and uncertainty characterization of the whole furnace temperature field, and provides reliable support for combustion optimization control and heating surface safety evaluation.

Description

Technical Field

[0001] This invention relates to the field of deep learning temperature field reconstruction technology, and in particular to a method for virtual-real fusion reconstruction of furnace temperature field based on digital twins. Background Technology

[0002] With the application of digital and intelligent technologies in the operation monitoring of coal-fired power plant boilers, some existing technologies attempt to introduce data-driven models or simulation analysis methods based on digital twins to intelligently assess and predict the furnace temperature field. However, existing intelligent technologies mostly rely on a limited number of furnace wall temperature measuring points or infrared surface measurement information, combined with empirical models or simple machine learning algorithms for inference, which is difficult to accurately reflect the complex three-dimensional temperature distribution characteristics inside the furnace. At the same time, some digital twin methods based on numerical simulation usually adopt offline calculation or static operating condition matching methods, which are difficult to achieve online adaptive calibration under the conditions of unit load fluctuations and frequent changes in combustion conditions. In addition, existing deep learning methods, when applied to furnace temperature field reconstruction, generally suffer from problems such as strong dependence on the scale of training data, easy drift of model parameters, and insufficient characterization of physical mechanisms and uncertainties, making it difficult to balance reconstruction accuracy, stability, and engineering usability, and still failing to meet the actual needs of coal-fired units for continuous, reliable, and interpretable reconstruction of the furnace temperature field. Summary of the Invention

[0003] To address the shortcomings of existing intelligent furnace temperature monitoring technologies, such as insufficient spatial coverage, difficulty in online correction of model biases, and limited ability to characterize uncertainties, this invention proposes a virtual-real fusion reconstruction method for the furnace temperature field based on digital twins. This method constructs a searchable prior benchmark for the temperature field using simulation results of the physical mechanisms of the furnace under multiple operating conditions. By introducing the unit's online operating conditions and measured wall temperature data into the state space model, the virtual temperature field is calibrated in real time, thereby achieving dynamic consistency constraints between the simulation model and the actual operating state. Building upon this, this invention further introduces a deep learning correction model for the three-dimensional temperature field and employs a low-rank adaptive parameter training mechanism with curvature-aware regularization. While maintaining the stability of the network backbone structure and pre-trained knowledge, this achieves efficient learning and precise correction of spatial biases in the furnace temperature under different operating conditions. Through the synergistic fusion of the aforementioned physical mechanism model, online observation data, and data-driven model, this invention not only achieves high-precision reconstruction of the entire furnace temperature field but also simultaneously outputs virtual measurement points and their uncertainty characterization results, providing a continuous, reliable, and interpretable temperature field sensing foundation for combustion optimization control and heating surface safety assessment.

[0004] This invention provides a method for reconstructing the virtual-real temperature field of a furnace based on digital twins, which specifically includes the following steps: Step S1: Digital Twin Model Construction: Obtain the three-dimensional structural drawings of the boiler furnace, the layout of the heating surface and the layout of the burner, construct a three-dimensional geometric model of the furnace including the furnace water-cooled wall, burner, screen superheater and reheater, and perform meshing of the three-dimensional geometric model of the furnace based on the furnace height and cross-section division rules to form a discrete mesh of the furnace. Step S2: Measurement point layout and data acquisition: Collect furnace operating condition data including furnace wall temperature data, flue gas temperature data, oxygen content data, unit load, primary air ratio, secondary air ratio, pulverized coal feed rate, and infrared images. Step S3: Solving the temperature field of physical mechanism: Extract the temperature deviation characteristics of each grid cell of the discrete grid of the furnace relative to the reference condition under different operating conditions, associate and store the temperature deviation characteristics with the corresponding operating condition label and grid number, and construct a thermal deviation benchmark database to characterize the spatial distribution variation law of furnace temperature. Step S4: Online calibration of virtual temperature field: During the actual operation of the unit, based on the current unit load, coal mill combination and air distribution method, the simulated temperature field closest to the current operating condition is retrieved from the thermal deviation benchmark database as the background virtual temperature field distribution data. A state space model is constructed with the background virtual temperature field distribution data as the state variable and the measured furnace wall temperature in the furnace operating condition data as the observation. Kalman filtering is used to use the measured furnace wall temperature as a hard constraint to calibrate and correct the deviation of the background virtual temperature field distribution data in real time, and the calibrated virtual temperature field distribution data is obtained. Step S5: Data-driven correction model construction: Pair the calibrated virtual temperature field distribution data with the furnace operating condition data at the corresponding time to construct a training sample set. Construct a 3D ResUNet on an industrial server, introduce a low-rank adaptive parameter training method based on curvature-aware regularization, and train the 3D ResUNet using the training sample set while keeping the 3D residual encoder-decoder network structure unchanged to obtain the trained 3D ResUNet. The temperature correction field covering the discrete grid of the furnace is obtained through the trained 3D ResUNet. According to the preset fusion strategy, the calibrated virtual temperature field distribution data and the temperature correction field are weighted and fused to generate a virtual-real fusion furnace temperature field reconstruction result.

[0005] Furthermore, the reconstructed furnace temperature field results, which integrate virtual and real data, are compared and displayed with the furnace surface temperature distribution output by the infrared thermal imaging camera. The internal temperature field of the furnace is displayed on the large screen in the central control room in the form of two-dimensional cross-section and three-dimensional cloud map, marking the high-temperature zone, low-temperature zone and local overheating area. The temperature field characteristic parameters and virtual measurement point data are sent to the combustion optimization control system to support the adjustment of combustion air distribution and the safety assessment of the heating surface.

[0006] Furthermore, a low-rank adaptive parameter training method based on curvature-aware regularization is introduced. While maintaining the 3D residual encoder-decoder network structure, the 3D ResUNet is trained using a training sample set to obtain the trained 3D ResUNet. The specific steps include: Step T1: Before training 3D ResUNet, a low-rank adaptive parameter structure is introduced into the trainable linear mapping layer of 3D ResUNet. The weight update of the trainable linear mapping layer is represented as a low-rank matrix product. LoRA parameter matrix A for low-rank projection and LoRA parameter matrix B for low-rank reconstruction are set as LoRA parameters. Step T2: Perform zero-effect initialization on LoRA parameter matrix A and LoRA parameter matrix B to ensure that the forward inference output of 3D ResUNet is consistent with its output in the pre-training state after the introduction of LoRA parameters. Through the above initialization method, the initial state of LoRA parameters is anchored near the pre-training solution. Step T3: Based on the training sample set, construct a proxy source domain sample set to approximate the distribution of the model pre-training data; during the construction process, the proxy source domain sample set is configured as a single proxy source domain sample set and multiple proxy source domain subsets to characterize the overall and distribution-level source domain features. Step T4: Perform Laplace uncertainty modeling under the condition of a single agent source domain sample set. Specifically, while keeping the parameters of the 3D ResUNet backbone network fixed and not participating in the update, use a single agent source domain sample set to perform a second-order Taylor expansion of the log-likelihood loss function corresponding to the LoRA parameters near the pre-trained solution, and construct the posterior probability distribution model of the LoRA parameters based on the second-order approximation result using the Laplace approximation method. Step T5: Use multiple surrogate source domain subsets to characterize the source domain feature distribution corresponding to the 3D ResUNet pre-training stage. As an extension of the Laplacian uncertainty modeling process, calculate the corresponding LoRA parameter precision matrix independently based on each surrogate source domain subset, and obtain the global LoRA parameter precision matrix by accumulating and fusing the precision matrices corresponding to each surrogate source domain subset. Step T6: Based on the global LoRA parameter accuracy matrix, construct a curvature-aware regularization structure to constrain the LoRA parameter update direction, and uniformly model the LoRA parameter update terms in the LoRA parameter parameter space to form the LaLoRA regularization function; Step T7: Introduce the LaLoRA regularization function into the training objective of 3D ResUNet to construct a joint optimization objective function with Laplacian regularization constraints. While keeping the 3D ResUNet backbone network structure and its pre-trained parameters fixed and not involved in updates, only perform parameter updates on LoRA parameter matrices A and B, and train the 3D ResUNet under the constraints of the joint optimization objective function through gradient optimization to obtain the 3D ResUNet trained with LaLoRA regularization constraints.

[0007] By adopting the above solution, the beneficial effects achieved by the present invention are as follows: Firstly, this invention, by introducing a multi-condition digital twin temperature field benchmark and an online state-space calibration mechanism, achieves a holistic reconstruction of the furnace temperature field from discrete measurement points to a continuous three-dimensional space. This enhances the spatial integrity and operational adaptability of the furnace's internal temperature distribution perception, solving the problem in existing technologies that rely solely on limited wall temperature measurement points or surface infrared information, making it difficult to reflect the true internal temperature field of the furnace. It also improves the reliability and engineering applicability of furnace temperature monitoring results under conditions of unit load fluctuations and frequent changes in combustion conditions. By using the simulated temperature field as a priori and combining it with measured wall temperatures for online calibration, this invention enables the virtual temperature field to be dynamically corrected according to the actual operating state of the unit, thus avoiding the problem of long-term deviation between the simple simulation model and actual operating conditions, and providing a stable foundation for the continuous online acquisition of the furnace temperature field.

[0008] Secondly, this invention introduces a deep learning correction model oriented towards the three-dimensional temperature field during the virtual-real fusion reconstruction of the furnace temperature field. Combined with a low-rank adaptive parameter training mechanism based on curvature-aware regularization, it achieves refined correction of spatial deviations in the temperature field without violating the original physical constraints. This improves the accuracy of the furnace temperature field reconstruction and the model's generalization ability, solving the problems of parameter drift and insufficient stability that easily occur in existing data-driven methods during operating condition migration and online fine-tuning. It also enhances the long-term consistency of the temperature field reconstruction results under multiple coal types and multiple load operating conditions. Through this training mechanism, this invention effectively suppresses disordered changes in the direction of key parameters while ensuring efficient and controllable model parameter updates, making the deep learning model more suitable for long-term online deployment and continuous operation in power plants.

[0009] Thirdly, by synchronously outputting the virtual measurement point temperature and its uncertainty characterization results from the virtual-real fusion temperature field, this invention expands the furnace temperature field reconstruction results from a single numerical output to a combination of "numerical value + confidence interval," enhancing the decision support capability of the temperature field results in combustion optimization control and heating surface safety assessment. It solves the problem in existing technologies where temperature field results lack confidence characterization and are difficult to use for risk assessment, thus strengthening the safety and engineering decision-making value of this invention in practical application scenarios such as combustion air distribution adjustment, local overheating early warning, and heating surface life assessment. Through the introduction of uncertainty information, this invention makes the furnace temperature field not only "visible" but also "reliable and usable," providing power plant operators with a more robust and interpretable basis for temperature field perception. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the process for a virtual-real fusion reconstruction method of furnace temperature field based on digital twin proposed in this invention; Figure 2 This is the average temperature distribution curve along the height of the furnace as proposed in Example 8. Detailed Implementation

[0011] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Example 1, according to Figure 1 , Figure 2 This invention provides a method for virtual-real fusion reconstruction of furnace temperature field based on digital twins, applied to the scenario of furnace temperature monitoring and combustion adjustment in coal-fired power plant boilers. The unit includes a furnace body mounted on the boiler steel frame, temperature measuring points arranged on the water-cooled walls around the furnace, infrared monitoring holes set on the side and rear walls of the furnace, and an industrial server arranged in the central control room. The method specifically includes the following steps: This embodiment uses a 600 MW subcritical coal-fired unit as the subject. The boiler is a π-type once-through boiler with an effective furnace height of approximately 30 m. Under stable operating conditions, the unit load is 580 MW, and four coal mills (A, B, C, and D) are in operation. The primary air to secondary air ratio is 38:62. The boiler burns medium-volatile bituminous coal with a lower heating value of approximately 19.5–21.0 MJ / kg. 48 sheathed thermocouple measuring points are arranged along the height of the furnace water-cooled wall, with 8 measuring points evenly distributed circumferentially in each layer, covering the four corner areas. Infrared thermal imaging cameras are installed at monitoring holes on the furnace sidewalls. The monitoring holes are equipped with cooling and purging devices to reduce the impact of high-temperature flue gas and ash on the imaging. An industrial server is installed in the central control room. The server has temperature field retrieval programs, online calibration programs, 3D reconstruction inference programs, and visualization display programs installed. It is connected to the boiler DCS system through redundant dual network ports to achieve real-time data exchange.

[0013] Step S1: Digital Twin Model Construction: Obtain the 3D structural drawings of the boiler furnace, the layout of the heating surface, and the layout of the burner. Construct a 3D geometric model of the furnace, including the furnace water-cooled wall, burner, screen superheater, and reheater. Based on the furnace height and cross-section division rules, the 3D geometric model of the furnace is meshed to form a discrete furnace mesh. In this embodiment, the meshing adopts a discrete strategy with 80 layers in the height direction (Nz) and 48 in the cross-section direction (Nx and Ny), so that the side length of the mesh unit voxels is in the range of 0.3 m to 0.6 m. A unique mesh number ID is assigned to each mesh unit, and a mesh index table is established. The mesh index table includes the mesh number, unit center coordinates, adjacency relationship, and heating surface area label, which are used for subsequent database indexing, measurement point mapping, and spatial alignment of reconstruction results. Step S2: Measurement Point Layout and Data Acquisition: Sheathed thermocouples are arranged at different heights on the furnace water-cooled wall. A flue gas temperature sensor and oxygen analyzer are arranged in the furnace outlet flue. An infrared thermal imaging camera equipped with cooling and purging devices is installed at the monitoring holes on the furnace sidewall. The thermocouples, flue gas temperature sensor, oxygen analyzer, and infrared thermal imaging camera are connected to the boiler DCS system via fieldbus. The DCS system then interacts with the industrial server via Ethernet to collect furnace operating condition data, including furnace wall temperature data, flue gas temperature data, oxygen data, unit load, primary air ratio, secondary air ratio, pulverized coal feed rate, and infrared images. In this embodiment, the sampling period for wall temperature and operating condition data is 1 second, and the infrared image acquisition frame rate is 10. Hz, and align all data with a unified timestamp to form a "slice of operating conditions at the same moment"; the oxygen content at the furnace outlet is stable at 3.6% to 4.2%, the flue gas temperature at the furnace outlet is in the range of 930℃ to 980℃, and the standard deviation of each wall temperature measuring point under short-term fluctuations is maintained in the range of 4℃ to 12℃. Step S3: Solving the physical mechanism temperature field: In an industrial server, a multi-condition three-dimensional CFD numerical simulation of the furnace is performed based on the Fluent platform. For each condition, a combustion-flow-heat transfer coupled numerical model is established on the discrete grid of the furnace. By solving the set of control equations, including the mass conservation equation, momentum conservation equation, energy conservation equation, and radiation heat transfer control equation, the three-dimensional temperature field distribution of the furnace under the corresponding condition is obtained. To cover common operating boundaries of the unit and improve the adaptability of the offline library to online conditions, the set of operating conditions constructed in this embodiment includes: unit loads of 450 MW, 520 MW, 580 MW, and 600 MW. The system covers four MW levels, with coal mill operation combinations including "ABCD", "ABC", and "BCD", and air distribution methods including secondary air enrichment, balanced air distribution, and primary air over-distribution. The three-dimensional temperature field distribution results of the furnace are uniformly formatted, binding the temperature field results for each operating condition to its corresponding unit load parameters, coal mill operation combination parameters, and air distribution method parameters. An index is established for the temperature values ​​of each grid unit according to the furnace discrete grid number, forming a multi-condition temperature field dataset that can be quickly retrieved based on operating conditions and spatial location. Based on this multi-condition temperature field dataset, statistical analysis is performed on the differences in temperature field distribution between different operating conditions. The temperature deviation characteristics of each grid unit in the furnace discrete grid relative to the reference operating condition under different operating conditions are extracted. These temperature deviation characteristics are associated and stored with the corresponding operating condition label and grid number, constructing a thermal deviation benchmark database to characterize the spatial distribution variation of furnace temperature. The database retrieval and update cycle is set to 10 seconds to balance online computational overhead and operating condition tracking sensitivity. Step S4: Online Calibration of Virtual Temperature Field: During actual unit operation, based on the current unit load, coal mill combination, and air distribution method, the simulated temperature field closest to the current operating conditions is retrieved from the thermal deviation benchmark database as the background virtual temperature field distribution data. A state-space model is constructed with the background virtual temperature field distribution data as the state variable and the measured furnace wall temperature from the furnace operating condition data as the observation. Kalman filtering is used to apply the measured furnace wall temperature as a hard constraint to the background virtual temperature field distribution data for real-time calibration and deviation correction, resulting in calibrated virtual temperature field distribution data. After online calibration, the calibration residual at the furnace wall temperature measurement point location decreases from 40℃~85℃ before calibration to the range of 8℃~20℃, and within 10... The temperature remains stable within the rolling window. Based on the calibrated temperature field, virtual measurement point temperature values ​​are output for the discrete grid positions of the furnace where no physical measurement points are arranged, and the uncertainty range of the virtual measurement points is given simultaneously to identify the difference in reconstruction reliability between sparse measurement point areas and obstructed field of view areas. In this embodiment, the uncertainty range near the burner layer and the center of the furnace is relatively small, typically ±25℃~±45℃, while the uncertainty range near the corner recirculation area and the area with strong obstruction of the heating surface increases, typically ±50℃~±80℃. Step S5: Data-Driven Correction Model Construction: The calibrated virtual temperature field distribution data and the corresponding furnace operating condition data are paired to construct a training sample set. A 3D ResUNet is built on an industrial server. A low-rank adaptive parameter training method based on curvature-aware regularization is introduced. While maintaining the 3D residual encoder-decoder network structure, the 3D ResUNet is trained using the training sample set to obtain the trained 3D ResUNet. ResUNet is used to obtain the temperature correction field covering the discrete grid of the furnace. According to the preset fusion strategy, the calibrated virtual temperature field distribution data and the temperature correction field are weighted and fused to generate the virtual-real fusion furnace temperature field reconstruction result. At the same time, the uncertainty field corresponding to the virtual measurement point is output, realizing the virtual-real fusion reconstruction and uncertainty characterization of the entire furnace temperature field. The construction method of the low-rank adaptive parameter training method based on curvature-aware regularization is as follows: Based on the fine-tuning mechanism of low-rank adaptive parameters, a parameter curvature-aware regularization structure based on Laplace uncertainty estimation is introduced to adaptively constrain the update amplitude of low-rank adaptive parameters in different curvature directions, thereby suppressing the drift of key parameter directions in the source domain while efficiently fine-tuning the parameters, thus constructing the low-rank adaptive parameter training method based on curvature-aware regularization. The reconstructed temperature field of the virtual-real fusion furnace was obtained, and the following characteristic parameters were calculated: (1) Average temperature distribution curve along the height of the furnace, vertical axis: furnace height (m), horizontal axis: average temperature (°C): Bottom of the furnace (0–5 m): average temperature approximately 860–890 ℃; In the middle of the furnace (12–18 m): the average temperature rises to 980–1000 ℃; Top of the furnace (25–30 m): Temperature drops to 900–930 °C.

[0014] (2) The area ratio of the high-temperature zone in the furnace cross-section and the location of the centroid of the high-temperature zone: The high-temperature region is defined as ≥1050 ℃, and statistics are performed on cross-sections at various heights: The high-temperature region in the central section accounts for the largest proportion of the area, approximately 28%–30%. The proportion of high-temperature area in the bottom and top regions of the furnace is less than 15%; The center of gravity of the high-temperature zone is located about 0.6 m off the secondary air side of the center of the furnace cross section.

[0015] (3) The highest temperature and duration in the area adjacent to the superheater and reheater: Based on the statistical data of the virtual-real fusion temperature field, as shown in Table 1: Table 1 .

[0016] (4) Results of combustion deviation judgment and results of overheating risk assessment of heating surface: Based on the results of the virtual-real furnace temperature field reconstruction, further calculations were performed: Average temperature difference between the left and right sides of the furnace: 42 ℃; The center of gravity of the high-temperature zone has long been shifted to the side where secondary winds are enriched. Judgment result: Moderate combustion deviation; The risk assessment results for the heated surface are as follows: Superheater area: Level II thermal stress risk; Reheater area: Level III local overheating risk (intervention required).

[0017] Example 2, based on Example 1, introduces a low-rank adaptive parameter training method based on curvature-aware regularization. While maintaining the 3D residual encoder-decoder network structure, the 3D ResUNet is trained using a training sample set to obtain the trained 3D ResUNet. The specific steps include: Step T1: Before training 3D ResUNet, a low-rank adaptive parameter structure is introduced into the trainable linear mapping layer of 3D ResUNet. The weight update of the trainable linear mapping layer is represented as a low-rank matrix product. LoRA parameter matrix A for low-rank projection and LoRA parameter matrix B for low-rank reconstruction are set as LoRA parameters. Step T2: Perform zero-effect initialization on LoRA parameter matrix A and LoRA parameter matrix B to ensure that the forward inference output of 3D ResUNet is consistent with its output in the pre-training state after the introduction of LoRA parameters. Through the above initialization method, the initial state of LoRA parameters is anchored near the pre-training solution, providing a unified reference benchmark for subsequent parameter uncertainty modeling and curvature estimation based on the pre-training solution, and avoiding abrupt changes in feature distribution due to parameter perturbation in the early stage of model training. Step T3: Based on the training sample set, construct a surrogate source domain sample set to approximate the distribution of the model's pre-training data. The surrogate source domain sample set is used to approximate the key feature patterns and parameter-sensitive directions learned by the model during the pre-training stage without accessing the original pre-training data. During the construction process, the surrogate source domain sample set is configured as a single surrogate source domain sample set and multiple surrogate source domain subsets to characterize the overall and distribution-level source domain features. By introducing a single surrogate source domain sample set and multiple surrogate source domain subsets, a data foundation is provided for estimating the loss curvature of LoRA parameters near the pre-training solution in subsequent steps, enabling parameter importance assessment to reflect the model's need to preserve source domain knowledge under single-source or multi-source conditions. Step T4: Perform Laplace uncertainty modeling under the condition of a single agent source domain sample set. Specifically, while keeping the parameters of the 3D ResUNet backbone network fixed and not participating in the update, use a single agent source domain sample set to perform a second-order Taylor expansion of the log-likelihood loss function corresponding to the LoRA parameters near the pre-trained solution. Based on the second-order approximation result, use the Laplace approximation method to construct the posterior probability distribution model of the LoRA parameters, thereby obtaining the accuracy matrix characterizing the uncertainty of the LoRA parameters. The formula for constructing the posterior probability distribution model of the LoRA parameters using the Laplace approximation method is as follows: ; in, This represents the amount of weight update introduced by LoRA at a certain linear mapping layer. This represents a single proxy source domain sample set. Indicates approximation. Indicates a sample set from a given proxy source domain. Under the condition of LoRA parameter increment The posterior probability distribution; Indicates the LoRA parameter increment Reference mean (center); Indicates the LoRA parameter increment The covariance matrix; Represents a normal distribution; Step T5: Use multiple surrogate source domain subsets to characterize the source domain feature distribution corresponding to the 3D ResUNet pre-training stage. As an extension of the Laplacian uncertainty modeling process, independently calculate the corresponding LoRA parameter accuracy matrix based on each surrogate source domain subset. Then, by accumulating and fusing the accuracy matrices corresponding to each surrogate source domain subset, obtain the global LoRA parameter accuracy matrix. The formula used is as follows: ; in, Represents the global precision matrix. Indicates the number of proxy source domain subsets. This represents the index of the proxy source domain subset. Indicates the first A proxy source domain subset dataset, Indicates based on the first Subdatasets of agent source domains The calculated LoRA parameter accuracy matrix; Step T6: Based on the global LoRA parameter accuracy matrix, construct a curvature-aware regularization structure to constrain the LoRA parameter update direction, and uniformly model the LoRA parameter update term in the LoRA parameter parameter space to form the LaLoRA regularization function, so as to limit the excessive offset of LoRA parameters in the direction of key parameters in the source domain during parameter fine-tuning, thereby achieving a balance between the target task learning ability and the source domain knowledge preservation ability. The curvature-aware regularization structure is a combination of parameter-wise quadratic penalty constraints based on the diagonal Fisher information matrix, intra-layer structured regularization constraints based on block diagonal Kronecker decomposition, and joint regularization constraints based on block tridiagonal Kronecker decomposition. 1. Based on the parameter-wise quadratic penalty constraint of the diagonal Fisher information matrix, a weighted quadratic penalty is applied to each LoRA parameter dimension to measure and constrain the source domain sensitivity in the single parameter dimension; 2. Based on the block diagonal Kronecker decomposition, the intra-layer structured regularization constraint introduces the accuracy approximation of the Kronecker structure within the LoRA adapter layer to characterize the correlation and cooperative variation relationship between the parameters within the same LoRA adapter; 3. Based on the joint regularization constraint of block tridiagonal Kronecker decomposition, the interaction terms between LoRA parameter matrix A and LoRA parameter matrix B are modeled simultaneously to jointly characterize the collaborative update relationship between low-rank projection and low-rank reconstruction parameters, thereby protecting the stability of the low-rank product structure in the key directions of the source domain during parameter fine-tuning. LaLoRA regularization function: ; in, Describes the regularization function for LaLoRA. This represents the offset of the LoRA parameter update relative to the reference solution; Indicates the transpose operation; Step T7: Introduce the LaLoRA regularization function into the training objective of 3D ResUNet to construct a joint optimization objective function with Laplacian regularization constraints. While keeping the 3D ResUNet backbone network structure and its pre-trained parameters fixed and not involved in updates, only perform parameter updates on LoRA parameter matrices A and B, and train the 3D ResUNet under the constraints of the joint optimization objective function through gradient optimization to obtain the 3D ResUNet trained with LaLoRA regularization constraints.

[0018] Example 3 differs from Example 2 in that the curvature-aware regularization structure in step T6 is different. Step T6 is different from that in Example 2. In this example, step T6 specifically includes the following: Step T6: Based on the global LoRA parameter accuracy matrix, a curvature-aware regularization structure is constructed, and the update terms of LoRA parameters are uniformly modeled in the parameter space of LoRA parameters to form the LaLoRA regularization function; the curvature-aware regularization structure is: parameter-wise quadratic penalty constraint based on the diagonal Fisher information matrix.

[0019] Example 4, based on Example 2, in this example, in step S2, the sheathed thermocouples on the furnace water-cooled wall are arranged in multiple layers along the height of the furnace, and several measuring points are arranged at intervals along the circumference of the furnace in each layer. Each measuring point is connected to the distributed I / O acquisition module through a protective junction box. The distributed I / O acquisition module is set in a local control box near the boiler body to reduce signal transmission interference and increase the sampling density of the furnace wall temperature space.

[0020] Example 5, based on Example 4, includes the following thermal deviation benchmark database: operating condition index information categorized by unit load, coal mill commissioning combination, and air distribution method; three-dimensional temperature field distribution data indexed by furnace discrete grid number under each operating condition; and three types of thermal deviation characteristic parameters: average temperature, maximum temperature, and temperature standard deviation, obtained based on furnace cross-section statistics. The thermal deviation benchmark database is used to provide an initial background virtual temperature field and thermal deviation reference for the furnace temperature field reconstruction process during the unit's online operation phase.

[0021] Example 6, based on Example 5, compares and displays the reconstructed furnace temperature field results of virtual-real fusion with the furnace surface temperature distribution output by the infrared thermal imaging camera. The internal temperature field of the furnace is displayed on the large screen in the central control room in the form of two-dimensional cross-section and three-dimensional cloud map, marking the high temperature zone, low temperature zone and local overheating area. The temperature field characteristic parameters and virtual measurement point data are sent to the combustion optimization control system to support combustion air distribution adjustment and heating surface safety assessment.

[0022] Example 7, based on Example 6, describes an industrial server equipped with a GPU accelerator card. This server is installed in the boiler control room cabinet and is used to carry the CFD offline simulation data management program and the 3D ResUNet inference program. It is connected to the DCS system through redundant dual network ports to realize the online calculation and real-time release of furnace temperature field reconstruction results.

[0023] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A method for reconstructing the virtual-real fusion of a furnace temperature field based on digital twins, characterized in that, The method specifically includes the following steps: Step S1: Obtain the three-dimensional structural drawings of the boiler furnace, the layout of the heating surface, and the layout of the burners. Construct a three-dimensional geometric model of the furnace and perform mesh generation to form a discrete mesh of the furnace. Step S2: Collect furnace operating condition data; Step S3: Extract the temperature deviation characteristics of the discrete grid in the furnace and construct a thermal deviation benchmark database; Step S4: Retrieve the simulated temperature field from the thermal deviation reference database as background virtual temperature field distribution data, and perform real-time calibration and deviation correction to obtain calibrated virtual temperature field distribution data; Step S5: Pair the calibrated virtual temperature field distribution data and furnace operating condition data to construct a training sample set; construct a 3D ResUNet, introduce a low-rank adaptive parameter training method based on curvature-aware regularization, train the 3D ResUNet using the training sample set to obtain a temperature-corrected 3D ResUNet; obtain the temperature correction field using the temperature-corrected 3D ResUNet; perform weighted fusion of the calibrated virtual temperature field distribution data and the temperature correction field to generate a virtual-real fusion furnace temperature field reconstruction result.

2. The method for reconstructing the virtual-real fusion of furnace temperature field based on digital twin according to claim 1, characterized in that: The construction method of the low-rank adaptive parameter training method based on curvature-aware regularization is as follows: Based on the fine-tuning mechanism of low-rank adaptive parameters, a parameter curvature-aware regularization structure based on Laplace uncertainty estimation is introduced to adaptively constrain the low-rank adaptive parameters, thus constructing the low-rank adaptive parameter training method based on curvature-aware regularization.

3. The method for reconstructing the virtual-real fusion of furnace temperature field based on digital twin according to claim 2, characterized in that: The training process of temperature-corrected 3D ResUNet includes the following steps: Step E1: In the trainable linear mapping layer of 3D ResUNet, set the LoRA parameter matrix A and LoRA parameter matrix B respectively, perform zero-effect initialization, and anchor to the vicinity of the pre-trained solution; Step E2: Based on the training sample set, construct a proxy source domain sample set, configured as a single proxy source domain sample set and multiple proxy source domain subsets; perform Laplace uncertainty modeling under the condition of a single proxy source domain sample set; use multiple proxy source domain subsets to characterize the source domain feature distribution in the 3D ResUNet pre-training stage, accumulate and fuse them to obtain the global LoRA parameter accuracy matrix. Step E3: Based on the global LoRA parameter accuracy matrix, construct a curvature-aware regularization structure to form the LaLoRA regularization function; introduce the LaLoRA regularization function into the training objective of 3D ResUNet, construct a joint optimization objective function, perform parameter updates, and obtain the temperature-corrected 3D ResUNet.

4. The method for reconstructing the virtual-real fusion of furnace temperature field based on digital twin according to claim 3, characterized in that: Laplace uncertainty modeling is performed by keeping the parameters of the 3D ResUNet backbone network fixed, using a single agent source domain sample set, performing a second-order Taylor expansion near the pre-trained solution, and using the Laplace approximation method to construct the posterior probability distribution model of the LoRA parameters.

5. The method for reconstructing the virtual-real fusion of furnace temperature field based on digital twin according to claim 3, characterized in that: Curvature-aware regularization structures are used to constrain the LoRA parameter update direction.

6. The method for reconstructing the virtual-real fusion of furnace temperature field based on digital twin according to claim 3, characterized in that: Curvature-aware regularization structures include at least one of the following forms or combinations thereof: parameter-wise quadratic penalty constraints based on diagonal Fisher information matrices; intra-layer structured regularization constraints based on block diagonal Kronecker decomposition; and joint regularization constraints based on block tridiagonal Kronecker decomposition.

7. The method for reconstructing the virtual-real fusion of furnace temperature field based on digital twin according to claim 3, characterized in that: The joint optimization objective function is subject to Laplace regularization constraints.

8. The method for reconstructing the virtual-real fusion of furnace temperature field based on digital twin according to claim 1, characterized in that: The industrial server is an industrial control server equipped with a graphics processing unit (GPU) accelerator card.

Images