Infrared temperature calibration method for positive electrode material sintering furnace based on space-frequency domain fusion denoising and emissivity prior guidance

Abstract

The application discloses an infrared temperature calibration method for positive electrode material sintering furnace based on space-frequency domain fusion denoising and emissivity prior guidance, which comprises the following steps: collecting an original infrared image and a corresponding original observation temperature matrix; inputting the original infrared image into a space-frequency domain feature fusion denoising network to obtain a denoised infrared image; inputting the original infrared image, the original observation temperature matrix and the denoised infrared image into an emissivity prior guided double-branch residual temperature calibration network to obtain a pixel-level emissivity prior graph, a global temperature compensation amount and a local space correction matrix; and superimposing the global temperature compensation amount and the local space correction matrix to the original observation temperature matrix to obtain a calibrated temperature field. The application can realize relatively accurate and stable infrared temperature calibration under the conditions of high temperature, strong noise and dynamic change of emissivity, and has good practical application effect.

Description

Technical Field

[0001] This invention relates to the field of infrared temperature measurement image processing, and in particular to an infrared temperature calibration method and system for cathode material sintering furnaces based on spatial-frequency domain fusion denoising and emission prior art guidance. Background Technology

[0002] Cathode materials are a key component of lithium-ion batteries, and their performance directly affects the battery's energy density, cycle life, and safety stability. In the cathode material preparation process, the sintering process is crucial in determining the material's crystal structure, particle morphology, and electrochemical performance. The uniformity and control precision of the temperature field distribution within the sintering furnace directly impact product quality consistency and batch stability. Therefore, accurate, continuous, and online monitoring of the temperature field inside the cathode material sintering furnace is of great significance for optimizing sintering process parameters, improving material preparation quality, and ensuring stable industrial production. Compared to traditional contact temperature measurement methods, infrared thermal imaging technology offers advantages such as non-contact operation, fast response speed, and the ability to acquire two-dimensional temperature field distributions, making it suitable for furnace temperature monitoring in high-temperature sintering environments.

[0003] However, infrared temperature detection still faces significant technical challenges in actual cathode material sintering furnace operations. On one hand, the complex high-temperature environment inside the furnace makes infrared images susceptible to factors such as detector thermal noise, stray radiation from the furnace chamber, atmospheric disturbances, and changes in radiation from high-temperature targets. This results in acquired infrared images with high noise levels, low signal-to-noise ratios, and blurred edges and thermal field details, thus affecting the accuracy of temperature field inversion. On the other hand, the surface state, particle structure, and radiation characteristics of cathode materials undergo dynamic changes during heating, isothermal processes, and sintering, causing the material's surface emissivity to become non-constant. Existing infrared temperature measurement methods typically rely on fixed emissivity or static calibration parameters for temperature conversion, which is ill-suited to emissivity drift and spatial non-uniformity errors during sintering, easily leading to significant deviations between infrared temperature measurements and actual temperatures. Therefore, how to simultaneously suppress dynamic noise and compensate for temperature measurement errors caused by emissivity changes under conditions of high noise and low-quality infrared images is a key problem that needs to be solved to achieve high-precision online temperature calibration of cathode material sintering furnaces.

[0004] Existing literature 1 [Wang M, Yuan P, Qiu S, et al. Dual-encoder UNet-based narrowband uncooled infrared imaging denoising network[J]. Sensors, 2025, 25(5): 1476.] proposes a narrowband uncooled infrared image denoising network based on dual-encoder U-Net. (See attached...) Figure 1As shown, this method addresses the problem of mixed Gaussian random noise and row / column fringe noise interference in narrowband uncooled infrared imaging. It constructs an infrared image denoising structure consisting of two encoder branches, a feature fusion module, and a decoder. One encoder branch employs a basic-scale residual attention structure to learn Gaussian random noise and local detail features; the other encoder branch employs an expanded-scale residual attention structure to learn large-scale noise features such as row / column fringe noise. Subsequently, the method fuses the noise features extracted from the two encoders at different scales and inputs the fused features into the decoder for image reconstruction, thereby suppressing mixed noise and improving image quality in narrowband infrared images.

[0005] Reference 2 [Chu N, Zheng S, Zhang S, et al. Temperature calibration of surface emissivities with an improved thermal image enhancement network[J]. Measurement, 2025: 118325.] proposes a method for temperature calibration of surface emissivities combined with thermal image enhancement. (See attached...) Figure 2 As shown, this method addresses the problem of inaccurate infrared temperature measurement results caused by differences in the emissivity of different material surfaces. First, it segments the input infrared sequence image to extract the target region and reference region. Then, based on the emissivity differences between different material regions, it normalizes the infrared image features of the target and reference regions. Finally, it integrates the feature information of the target and reference regions through an image enhancement network with a cross-attention mechanism, aligning the grayscale distribution of the target region with the statistical characteristics of the reference region, such as mean, variance, and probability distribution. Based on the calibration relationship between grayscale values ​​and actual temperature, the enhanced infrared image is converted into a temperature curve or temperature distribution, thereby correcting the infrared temperature measurement deviation caused by surface emissivity differences.

[0006] Reference 1 can improve the signal-to-noise ratio (SNR) of infrared images, but it mainly targets Gaussian random noise and row and column fringe noise in uncooled infrared imaging. Its goal is to improve the visual quality and SNR of infrared images, but it does not specifically model the non-stationary and time-varying characteristics of noise in cathode material sintering furnaces as a function of temperature, furnace radiation environment, and sintering stage, nor does it consider the impact of denoising results on the accuracy of subsequent infrared temperature measurement. Reference 2 can compensate for the impact of different emissivity differences of different materials on temperature measurement results to some extent, but it mainly relies on the segmentation of the target area and the reference area, statistical alignment, and gray-scale-temperature mapping relationship to compensate for the temperature measurement deviation caused by different emissivity differences of different materials. It does not set up a dedicated front-end denoising structure for problems such as strong thermal noise, low SNR, and blurred thermal field boundaries in infrared images of industrial sintering furnaces, nor does it form a unified processing flow from noisy infrared image denoising to emission prior estimation and then to temperature field calibration. In actual sintering processes, the infrared images inside the furnace are not only subject to dynamic noise interference, but the surface condition and emissivity of the material also continuously change over time. Simple image denoising or separate emissivity calibration methods are insufficient to meet the requirements for high-precision online temperature measurement. Therefore, existing technologies still struggle to achieve coordinated processing from noisy infrared image acquisition, dynamic noise suppression, emissivity prior estimation to temperature field calibration and reconstruction. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] Based on the above, the purpose of this invention is to provide an infrared temperature calibration method for cathode material sintering furnaces based on spatial-frequency domain fusion denoising and emission prior guidance, in order to solve the technical problem that existing infrared temperature measurement methods are easily affected by strong noise interference and emissivity changes in the high-temperature dynamic sintering environment of cathode material sintering furnaces, resulting in insufficient temperature measurement accuracy.

[0009] (II) Technical Solution

[0010] This invention discloses an infrared temperature calibration method for cathode material sintering furnaces based on spatial-frequency domain fusion denoising and emission prior art guidance, characterized by comprising the following steps:

[0011] Acquire raw infrared images of the cathode material sintering furnace I noisy and its corresponding original observed temperature matrix T raw ;

[0012] The original infrared image I noisy The input is fed into a spatial-frequency domain feature fusion denoising network trained using a blind spot masking self-supervised strategy to obtain a denoised infrared image I. denoised ;

[0013] The original infrared image I noisy Original observed temperature matrix T rawand denoised infrared images I denoised The data is input to a launch-first a priori-guided bi-branch residual temperature calibration network, which is jointly trained based on physical constraints and sparse thermocouple measurement point supervision, and includes a launch-first a priori estimation module and a bi-branch residual temperature calibration module.

[0014] Based on the original infrared image I, the emission prior a priori estimation module is used. noisy and the original observed temperature matrix T raw Feature extraction is performed to obtain a pixel-level emission prior image E;

[0015] The denoised infrared image I denoised Launch first-pass diagram E and original observation temperature matrix T raw The input to the dual-branch residual temperature calibration module is used for channel splicing and feature encoding, and the global temperature compensation amount Δ is output through the global temperature compensation branch. g The local space correction matrix Δ is output through the local space correction branch. l ;

[0016] The global temperature compensation amount Δ g and the local space correction matrix Δ l Superimposed on the original observed temperature matrix T raw The calibrated temperature field T is obtained. perd .

[0017] (III) Beneficial Effects

[0018] 1. This invention uses a spatial-frequency domain feature fusion denoising network to effectively suppress random noise, local artifacts and high-frequency disturbances in infrared images under high-temperature dynamic conditions in cathode material sintering furnaces. It also preserves information on the furnace's thermal field boundaries, temperature gradients and thermal transition regions during the denoising process, avoiding the loss of thermal field details caused by the denoising process. This makes the denoised infrared images more suitable for subsequent temperature inversion and calibration.

[0019] 2. This invention utilizes a dual-branch residual temperature calibration network guided by the emission first a priori map to compensate for the effects of changes in material surface emissivity, stray radiation within the furnace, and spatial non-uniformity on infrared temperature measurement results during the sintering process of the cathode material. Through emission first a priori estimation, global temperature compensation, and local spatial correction, it can simultaneously correct systematic temperature deviations and local temperature measurement errors, thereby improving the calibration accuracy of the temperature field within the furnace.

[0020] 3. This invention does not rely on complete noise-free infrared image labels and dense temperature true value labels. It completes model optimization through blind spot self-supervised training, infrared radiation physical constraints, and supervision by a small number of thermocouple measurement points. It is suitable for online infrared temperature monitoring of cathode material sintering furnaces under conditions of high temperature, strong noise, dynamic emissivity changes, and scarce temperature labels. It significantly reduces the data labeling cost in industrial fields and has extremely high industrial deployment applicability and robustness.

[0021] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0023] Figure 1 This is a schematic diagram of the infrared image denoising network structure based on dual encoder U-Net in the background art of this invention;

[0024] Figure 2 This is a schematic diagram of the surface emissivity temperature calibration method based on thermal image enhancement network in Reference 2 of the background art of this invention;

[0025] Figure 3 This is a flowchart of the overall process of the infrared temperature calibration method based on spatial-frequency domain fusion denoising and emission prior art guided by a preferred embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the spatial frequency domain feature fusion denoising network structure of a preferred embodiment of the present invention;

[0027] Figure 5 This is a schematic diagram of the dual-branch residual temperature calibration network structure guided by launch prior a priori in a preferred embodiment of the present invention;

[0028] Figure 6 This is a denoised infrared image sequence of the sintering process according to a preferred embodiment of the present invention;

[0029] Figure 7 This is a corrected temperature distribution thermogram of the sintering process according to a preferred embodiment of the present invention. Detailed Implementation

[0030] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0031] Furthermore, unless otherwise defined, the technical or scientific terms used in this application description shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "upper," "lower," "left," "right," "center," "vertical," "horizontal," "inner," and "outer," etc., used in this application description to indicate relative direction or positional relationship are used only to indicate relative orientation or positional relationship, and do not imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. When the absolute position of the described object changes, its relative positional relationship may also change accordingly, and therefore should not be construed as a limitation on this application. The terms "first," "second," "third," and similar terms used in this application description are used only for descriptive purposes to distinguish different components, and should not be construed as indicating or implying relative importance. The terms "a," "one," or "the," etc., used in this application description should not be construed as an absolute limitation on quantity, but should be construed as indicating the existence of at least one. The terms "including," "comprising," etc., used in this application description mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, without excluding other elements or objects.

[0032] It should also be noted that, unless otherwise explicitly specified and limited, the terms such as “installation,” “connection,” and “linkage” used in the description of this application should be interpreted broadly. For example, a connection can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can also refer to the internal connection of two components. Those skilled in the art can understand its specific meaning in this application according to the specific circumstances.

[0033] like Figure 3 As shown in this embodiment, an infrared temperature calibration method for cathode material sintering furnaces based on spatial-frequency domain fusion denoising and emission prior art guidance includes the following steps:

[0034] Acquire raw infrared images of the cathode material sintering furnace I noisy and its corresponding original observed temperature matrix T raw ;

[0035] The original infrared image I noisy The input is fed into a spatial-frequency domain feature fusion denoising network trained using a blind spot masking self-supervised strategy to obtain a denoised infrared image I. denoised ;

[0036] The original infrared image I noisy Original observed temperature matrix T raw and denoised infrared images I denoisedThe data is input to a launch-first a priori-guided bi-branch residual temperature calibration network, which is jointly trained based on physical constraints and sparse thermocouple measurement point supervision, and includes a launch-first a priori estimation module and a bi-branch residual temperature calibration module.

[0037] Based on the original infrared image I, the emission prior a priori estimation module is used. noisy and the original observed temperature matrix T raw Feature extraction is performed to obtain a pixel-level emission prior image E;

[0038] The denoised infrared image I denoised Launch first-pass diagram E and original observation temperature matrix T raw The input to the dual-branch residual temperature calibration module is used for channel splicing and feature encoding, and the global temperature compensation amount Δ is output through the global temperature compensation branch. g The local space correction matrix Δ is output through the local space correction branch. l ;

[0039] The global temperature compensation amount Δ g and the local space correction matrix Δ l Superimposed on the original observed temperature matrix T raw The calibrated temperature field T is obtained. perd .

[0040] The method in this embodiment uses an infrared high-temperature endoscopic detector to acquire the original infrared image I inside the cathode material sintering furnace. noisy and the original observed temperature matrix T raw As input, the original infrared image I is first processed by a spatial-frequency domain feature fusion denoising network. noisy Dynamic noise in the process is suppressed, and then a dual-branch residual temperature calibration network guided by the transmit prior apocalypse is used to compensate for systematic temperature deviations and local spatial non-uniform errors in infrared thermometry. Finally, the calibrated furnace temperature field T is output. perd .

[0041] Among them, such as Figure 4 As shown, the spatial-frequency domain feature fusion denoising network adopts an encoder-decoder structure. A spatial branch, a frequency branch, and a spatial-frequency cross-attention module are provided in the bottleneck layer between the encoder and decoder. The spatial branch is used to extract thermal field context information, temperature gradient information, and thermal transition region features within the sintering furnace. The frequency branch is used to decompose deep features into the frequency domain and suppress high-frequency noise. The spatial-frequency cross-attention module guides the selection of frequency domain details through spatial features and supplements the boundary and texture information in the spatial features with frequency features. It then performs feature fusion on the spatial and frequency features and outputs the fused spatial-frequency domain features.

[0042] Specifically, the encoder extracts multi-scale thermal field features from the raw infrared image, the decoder restores the image spatial resolution and outputs a denoised infrared image, the spatial branch extracts furnace thermal field context information, temperature gradient information, and thermal transition region features, and the frequency branch decomposes deep features into the frequency domain and suppresses high-frequency noise. Subsequently, the spatial and frequency domain features are fused through a spatial-frequency cross-attention module, and an adaptive weighted fusion module controls the transmission ratio of shallow detail features and deep semantic features during the decoding stage, thereby reducing noise propagation to the decoding end while preserving the details of the thermal field structure.

[0043] Let F be the deep feature output by the encoder. L The frequency domain branch is first decomposed by discrete wavelet transform as shown in equation (1):

[0044] (1)

[0045] in, Represents discrete wavelet transform. This represents the low-frequency approximate subband, used to characterize the overall thermal field structure and regional temperature hierarchy information; , , These represent high-frequency detail subbands in different directions, used to characterize edge information, local gradient changes, and high-frequency noise perturbations.

[0046] For the high-frequency subbands obtained by decomposition, a dynamic spectrum gating module is used for feature modulation and noise suppression, and the frequency domain enhanced features are reconstructed by inverse discrete wavelet transform as shown in equation (2):

[0047] (2)

[0048] in, This indicates a dynamic spectrum gating module. Represents the inverse discrete wavelet transform. This represents the frequency domain enhancement characteristics. The dynamic spectrum gating module generates adaptive gating weights based on the channel statistical characteristics of the high-frequency subbands, suppressing the high-frequency response with strong noise components while retaining effective detailed information related to thermal field boundaries and temperature abrupt regions.

[0049] The spatial branch uses a multi-scale dilated convolutional structure to extract spatial context features, and its feature extraction process is shown in equation (3):

[0050] (3)

[0051] in, This represents a pyramid pooling module for empty spaces. This represents multi-scale features in the spatial domain. By setting dilated convolution branches with different expansion rates, the spatial branch can simultaneously acquire local thermal details, thermal transition region boundaries, and large-scale thermal field distribution information, thereby enhancing the network's ability to perceive the complex spatial structure of the temperature field inside the furnace.

[0052] After obtaining frequency domain enhancement features and spatial multi-scale features Then, the dual-domain features are fused using the space-frequency cross-attention module as shown in equation (4):

[0053] (4)

[0054] in, This indicates the space-frequency cross-attention module. This represents the fused spatial-frequency domain features. The spatial-frequency cross-attention module guides the selection of frequency domain details through spatial features and supplements the boundary and texture information in the spatial features with the frequency domain features, enabling the network to balance noise suppression and thermal field structure preservation under strong noise conditions.

[0055] During the decoding stage, the fused spatial-frequency domain features are upsampled layer by layer and fused with the shallow features of the encoder. To avoid noise in the shallow features being directly transmitted to the decoding end, an adaptive weighted fusion module is used to weight the shallow and deep features, and the fused features for the decoding stage are obtained through equation (5). :

[0056] (5)

[0057] in, This indicates the adaptive weighted fusion module. This represents the shallow features output by the i-th layer of the encoder. Indicates the decoder's first Deep features of the layer Indicates an upsampling operation. and This represents the learnable weights. This module allows for dynamic adjustment of the contribution of features at different levels to the denoising result, preventing shallow noise from being directly transmitted to the output and improving the quality of the denoised image. denoised Fidelity of the thermal field boundary, temperature gradient, and regional structure.

[0058] like Figure 5As shown, the launch-a priori-guided bi-branch residual temperature calibration network includes a launch-a priori estimation module and a bi-branch residual temperature calibration module. The launch-a priori estimation module estimates the pixel-level launch-a priori map based on the original infrared observation information; the bi-branch residual temperature calibration module calculates the global temperature compensation and local spatial correction matrix based on the denoised infrared image, the launch-a priori map, and the original observation temperature matrix, and outputs the final calibrated temperature field T. perd .

[0059] The transmitter first a priori estimation module receives the raw infrared image I noisy and the original observed temperature matrix T raw And by splicing them along the channel dimension, the launch prior art map E is obtained through equation (6).

[0060] (6)

[0061] in, This indicates a channel splicing operation. This indicates the launch prior a priori estimation module, which uses a U-net architecture. The parameter represents the learnable parameter of this module, and E represents the estimated pixel-level emission prior map. This emission prior map E is used to characterize the spatial difference in emissivity of the cathode material during sintering due to changes in surface state, particle structure, and radiation characteristics, providing physical prior information for subsequent temperature calibration.

[0062] The dual-branch residual temperature calibration module receives denoised infrared images. denoised Launch first-pass diagram E and original observation temperature matrix T raw The three data points are then concatenated and feature-encoded to obtain the high-level thermal field features. Based on these features, the network is divided into a global temperature compensation branch and a local spatial correction branch. The global temperature compensation branch outputs the global temperature compensation amount Δ. g Used to correct systematic biases in infrared thermometry; the local space correction branch outputs the local space correction matrix Δ. l This is used to compensate for non-uniform temperature measurement errors at different spatial locations within the furnace. The process is shown in equation (7):

[0063] (7)

[0064] in, This indicates a dual-branch residual temperature calibration module. This indicates the learnable parameters of the module. This represents the global temperature compensation amount. This represents the local space correction matrix.

[0065] Finally, the global temperature compensation amount and local space correction moments Array superimposed on the original observed temperature matrix The calibration temperature field is obtained through equation (8):

[0066] (8)

[0067] in, This represents the final output calibration temperature field. This represents a matrix of all ones, used to represent the global temperature compensation amount. The signal is broadcast throughout the entire image space. This method simultaneously compensates for both the overall offset error and spatial non-uniformity error during infrared thermometry, thereby improving the accuracy of temperature field detection within the cathode material sintering furnace.

[0068] In the model training process, this embodiment adopts a phased training method. The first phase is the self-supervised training phase of the spatial-frequency domain feature fusion denoising network. A blind spot masking strategy is used to randomly mask the original noisy infrared image, so that the network can predict the masked pixel based on the neighboring pixels around the target pixel, thereby learning the noiseless radiation representation in the infrared image. Its denoising loss is shown in Equation (9):

[0069] (9)

[0070] in, This represents the blind spot self-supervised denoising loss. This represents the set of pixels selected for the blind spot mask. This represents any one of the occluded pixels. This represents a small constant used to improve the stability of the loss function. This loss function allows for the training of denoising networks even when clean infrared image labels are unavailable.

[0071] The second stage is the training stage of the launch-first a priori-guided dual-branch residual temperature calibration network. This stage employs a combination of physical constraints and sparse measurement point supervision for optimization. The physical constraints, based on the Stefan-Boltzmann law, are used to constrain the consistency between emissivity, radiation intensity, and temperature; the sparse measurement point supervision is provided by thermocouple temperature measurements within the furnace, used to correct for absolute temperature deviations in the calibration temperature field.

[0072] Launch Priority Physical Consistency Loss L SB-e As shown in equation (10):

[0073] (10)

[0074] in, This indicates that the launch first verified the physical consistency loss. Represents the Stefan-Boltzmann constant. This indicates element-wise multiplication. This represents the scaling normalization operator. This loss is used to constrain the spatial consistency between the estimated emission prior map and the original infrared radiation distribution.

[0075] Radiation uniformity loss L during temperature reconstruction phase SB-t As shown in equation (11):

[0076] (11)

[0077] in, This represents the radiation consistency loss during the temperature reconstruction stage. This loss is used to constrain the consistency between the calibrated temperature field and the radiation distribution corresponding to the denoised infrared image, ensuring that the output temperature field conforms to the physical laws of infrared radiometric thermometry.

[0078] The monitoring loss of sparse thermocouple measuring points is shown in equation (12):

[0079] (12)

[0080] in, This indicates the monitoring loss at sparse thermocouple measuring points. Indicates the first The mapping position of each thermocouple measuring point in the infrared image coordinate system This indicates the actual temperature reading at the thermocouple measuring point. This indicates the number of thermocouple measuring points. This loss allows for the constraint of the overall temperature calibration results using a limited number of contact temperature measuring points.

[0081] The total loss function of the launch-first a priori guided dual-branch residual temperature calibration network is shown in equation (13):

[0082] (13)

[0083] in, This represents the total loss during the training phase of the temperature calibration network. , , This represents the weighting coefficients corresponding to each loss term. Using the above loss function, temperature field calibration can be completed by leveraging infrared radiation physical constraints and a small number of thermocouple measurement points, even in the absence of full-field real temperature labels.

[0084] Using the above method, infrared image sequences acquired during the heating and isothermal stages of an actual sintering process in a cathode material sintering furnace are processed to obtain denoised infrared image sequences and calibrated temperature distribution thermogram sequences at different temperatures, as shown below. Figure 6 and Figure 7 As shown. Among them, Figure 6 The results of the present invention on the original infrared image after spatial-frequency domain fusion denoising are shown. It can be seen that random noise and local artifacts in the denoised infrared image are significantly suppressed, and the thermal field boundary, thermal transition region and main temperature distribution structure in the furnace are well preserved. Figure 7 The temperature distribution thermogram obtained by the dual-branch residual temperature calibration network based on launch prior a priori guidance is shown. It can be seen that the calibrated temperature field distribution is more continuous, local abnormal temperature fluctuations are corrected, and it can reflect the temperature changes in different areas of the sintering furnace more stably.

[0085] Furthermore, the calibration temperature field T perd The predicted values ​​at key temperature measurement points were compared with the actual measured temperatures of the thermocouples, and the mean absolute error (MAE) and root mean square error (RMSE) were calculated. The results are shown in Table 1. It can be seen that the temperature measurement error at key points after correction is significantly lower than that of the original temperature matrix at key points. This indicates that the present invention improves the temperature prediction accuracy at key temperature measurement points, demonstrating that the present invention can achieve relatively accurate and stable infrared temperature calibration under conditions of high temperature, strong noise, and dynamic changes in emissivity, and has good practical application effects.

[0086] Table 1 Temperature measurement errors at key points during sintering

[0087]

[0088] In summary, this invention suppresses dynamic noise in sintering furnace infrared images through a spatial-frequency domain feature fusion denoising network, and achieves synergistic compensation for systematic temperature measurement deviations and local spatial non-uniform errors through an emission prior a priori-guided bi-branch residual temperature calibration network, thereby completing the conversion from the original noisy infrared image to a high-precision calibrated temperature field. This method does not require complete, high-quality, noise-free infrared image tags, nor does it require densely packed temperature true value tags across the entire field. It is suitable for online infrared temperature detection and calibration under conditions of high temperature, strong noise, dynamic emissivity changes, and scarce temperature tags in cathode material sintering furnaces.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for infrared temperature calibration of cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance, characterized in that, Includes the following steps: Acquire raw infrared images of the cathode material sintering furnace I noisy and its corresponding original observed temperature matrix T raw ; The original infrared image I noisy The input is fed into a spatial-frequency domain feature fusion denoising network trained using a blind spot masking self-supervised strategy to obtain a denoised infrared image I. denoised ; The original infrared image I noisy Original observed temperature matrix T raw and denoised infrared images I denoised The data is input to a launch-first a priori-guided bi-branch residual temperature calibration network, which is jointly trained based on physical constraints and sparse thermocouple measurement point supervision, and includes a launch-first a priori estimation module and a bi-branch residual temperature calibration module. Based on the original infrared image I, the emission prior a priori estimation module is used. noisy and the original observed temperature matrix T raw Feature extraction is performed to obtain a pixel-level emission prior image E; The denoised infrared image I denoised Launch first-pass diagram E and original observation temperature matrix T raw The input to the dual-branch residual temperature calibration module is used for channel splicing and feature encoding, and the global temperature compensation amount Δ is output through the global temperature compensation branch. g The local space correction matrix Δ is output through the local space correction branch. l ; The global temperature compensation amount Δ g and the local space correction matrix Δ l Superimposed on the original observed temperature matrix T raw The calibrated temperature field T is obtained. perd .

2. The infrared temperature calibration method for cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance according to claim 1, characterized in that, The spatial-frequency domain feature fusion denoising network adopts an encoder-decoder structure, with a spatial branch, a frequency branch, and a spatial-frequency cross-attention module in the bottleneck layer between the encoder and the decoder. The spatial branch is used to extract thermal field context information, temperature gradient information and thermal transition region features in the sintering furnace, and the frequency domain branch is used to decompose deep features into the frequency domain and suppress high-frequency noise. The spatial-frequency cross-attention module guides the selection of frequency domain details through spatial domain features, and supplements the boundary and texture information in the spatial domain features with frequency domain features. It then performs feature fusion on the spatial and frequency domain features and outputs the fused spatial-frequency domain features.

3. The infrared temperature calibration method for cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance according to claim 2, characterized in that, The frequency domain branch affects deep feature F L Perform discrete wavelet transform to decompose the low-frequency approximate subband. And the high-frequency detail subbands in different directions, the decomposition formula is: in, Represents discrete wavelet transform. Indicates a low-frequency approximate subband; , , These represent high-frequency detail subbands in different directions; The frequency domain branch utilizes a dynamic spectrum gating module to generate adaptive gating weights based on the channel statistical characteristics of the high-frequency detail subbands. This suppresses high-frequency responses with strong noise components while retaining effective detail information related to thermal field boundaries and temperature abrupt regions. The frequency domain enhancement features are then reconstructed through inverse transform, with the reconstruction formula as follows: in, This indicates a dynamic spectrum gating module. Represents the inverse discrete wavelet transform. This indicates frequency domain enhancement features.

4. The infrared temperature calibration method for cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance according to claim 2, characterized in that, During the upsampling process of the decoder, the shallow features output by the encoder and the corresponding deep features of the decoder are weighted by an adaptive weighted fusion module to obtain the fused features for the decoding stage. : in, This indicates the adaptive weighted fusion module. This represents the shallow features output by the i-th layer of the encoder. This represents the deep features of the i-th layer of the decoder. Indicates an upsampling operation. and Indicates learnable weights; The adaptive weighted fusion module dynamically adjusts the contributions of the shallow and deep features to the denoising result using learnable weights, thereby improving the denoising performance of the infrared image I. denoised Fidelity of the thermal field boundary, temperature gradient, and regional structure.

5. The infrared temperature calibration method for cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance according to claim 1, characterized in that, The dual-branch residual temperature calibration module includes a feature encoding unit, a global temperature compensation branch, and a local spatial correction branch; The feature encoding unit is used to process the stitched denoised infrared image I denoised Launch first-pass diagram E and original observation temperature matrix T raw Feature encoding is performed to generate high-level thermal field features; The global temperature compensation branch is used to output a single scalar global temperature compensation amount Δ based on the characteristics of the high-level thermal field. g To correct systematic biases in infrared thermometry; The local space correction branch is used to output a two-dimensional local space correction matrix Δ based on the high-level thermal field characteristics. l This is to compensate for the non-uniform temperature measurement error at different spatial locations within the sintering furnace.

6. The infrared temperature calibration method for cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance according to claim 1, characterized in that, The calibrated temperature field The calculation formula is: in, This represents a matrix of all ones, used to represent the global temperature compensation amount. Broadcast to the entire image space.

7. The infrared temperature calibration method for cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance according to claim 1, characterized in that, The denoising loss function of the spatial-frequency domain feature fusion denoising network trained using a blind spot self-supervised approach is as follows: in, This represents the blind spot self-supervised denoising loss. This represents the set of pixels selected for the blind spot mask. This represents any one of the occluded pixels. This represents a small constant used to improve the stability of the loss function.

8. The infrared temperature calibration method for cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance according to claim 1, characterized in that, During the training process, the launch-first a priori-guided dual-branch residual temperature calibration network, whose physical constraints are based on the Stefan-Boltzmann law, suffers a physical consistency loss L. SB include: Launch Priority Physical Consistency Loss L SB-e This is used to constrain the spatial consistency between the estimated emission prior map E and the original infrared radiation distribution; And, the radiation uniformity loss L during the temperature reconstruction phase SB-t Used to constrain the calibrated temperature field T perd With denoised infrared image I denoised The physical consistency between the corresponding radiation distributions.

9. The infrared temperature calibration method for cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance according to claim 8, characterized in that, During the training process of the launch-first a priori guided dual-branch residual temperature calibration network, the supervision loss of the sparse thermocouple measuring points is: in, This indicates the monitoring loss at sparse thermocouple measuring points. Indicates the first The mapping position of each thermocouple measuring point in the infrared image coordinate system This indicates the actual temperature reading at the thermocouple measuring point. This indicates the number of thermocouple measuring points.

10. The infrared temperature calibration method for cathode material sintering furnace based on spatial-frequency domain fusion denoising and emission prior art guidance according to claim 8, characterized in that, The total loss function of the launch prior art-guided dual-branch residual temperature calibration network is: in, This represents the total loss during the training phase of the temperature calibration network. , , This represents the weighting coefficient corresponding to each loss term.

Images