A whiteness closed-loop detection and quality tracing system on a flake caustic soda production line

Abstract

The present application relates to the field of flake caustic soda production, and discloses a whiteness closed-loop detection and quality tracing system for a flake caustic soda production line, comprising: an image acquisition module for synchronously acquiring a visible light multi-angle polarization image of flake caustic soda on a flake caustic soda conveying line; a radiation correction module for obtaining a temperature correction image; a coupling correction module for establishing a joint observation equation and solving to separate out a diffuse reflection image; a adhesion segmentation module for image segmentation of the diffuse reflection image; a detection module for outputting the whiteness value and corresponding confidence of each flake caustic soda; a feedback module for closed-loop control of whiteness; and a quality tracing module for locating an abnormal period when quality is abnormal. By synchronously acquiring multispectral and polarization images and performing multi-stage correction, a pure diffuse reflection image is obtained; based on a multi-task deep learning network, the whiteness value and confidence of each flake caustic soda are output, so that online accurate detection and real-time closed-loop regulation and control of flake caustic soda whiteness are realized.

Description

Technical Field

[0001] This invention relates to the field of caustic soda production technology, specifically to a closed-loop detection and quality traceability system for whiteness on a caustic soda production line. Background Technology

[0002] Sodium hydroxide flakes are a basic chemical raw material widely used in industries such as papermaking, textiles, and water treatment. The whiteness of sodium hydroxide flakes is one of the key indicators for measuring its purity and appearance quality. The whiteness value directly reflects the amount of impurities in the product and is an important basis for product grading and pricing. In the continuous production process of sodium hydroxide flakes, how to detect the whiteness of the product in real time and accurately, and adjust the process parameters in a timely manner based on the test results, is of great significance for stabilizing product quality and reducing production costs.

[0003] Currently, the whiteness detection on caustic soda production lines mainly involves periodically sampling from the end of the conveyor belt or the packaging line and using a blue light whiteness meter for testing. However, after caustic soda is fed from the rotary drum scraper, it is in the early stage of forming on the conveyor belt, and its physical state is extremely unstable. The temperature of freshly scraped caustic soda can reach over 80 degrees Celsius. According to the blackbody radiation law, the thermal radiation generated by a high-temperature object will be superimposed on the visible light image, causing the overall brightness of the image acquired by the visual inspection system to be too high, resulting in a positive deviation in the whiteness measurement value. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a closed-loop detection and quality traceability system for whiteness in caustic soda flake production lines, solving the problem of positive deviations in whiteness detection on existing caustic soda flake production lines.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a closed-loop detection and quality traceability system for whiteness on a caustic soda flake production line, comprising: The image acquisition module is used to simultaneously acquire visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images of caustic soda flakes on the caustic soda flake conveying line, and perform geometric registration. The radiation correction module is used to invert the true temperature distribution of the caustic soda surface based on the near-infrared multispectral image and the thermal infrared image, calculate the thermal radiation-added gray value and subtract it from the visible light multi-angle polarization image to obtain a temperature-corrected image. The coupling correction module is used to combine the temperature-corrected image with the visible light multi-angle polarization image, establish a joint observation equation, and solve to separate the diffuse reflection image. The adhesion segmentation module is used to segment the diffuse reflection image, construct shape priors based on a statistical shape model, and extract the mask for each individual caustic soda flake. The detection module is used to input the visible light multi-angle polarization image, temperature-corrected image, diffuse reflection image and the mask into a pre-trained multi-task deep learning network, output the whiteness value and corresponding confidence level of each caustic soda flake, and calculate the overall whiteness detection result at the current time based on the weighted average of the confidence level. The feedback module is used to filter the whiteness value and adjust the sampling frequency according to the confidence level, and adjust the parameters of the caustic soda production line according to the filtered whiteness value to perform closed-loop control of whiteness. The quality traceability module is used to timestamp and store the overall whiteness test results, confidence levels, and corresponding caustic soda production line parameters at each moment, and to locate the abnormal period when quality is abnormal.

[0006] By adopting the above technical solution, multispectral and polarization images are acquired simultaneously and corrected at multiple levels to eliminate the dynamic interference of high temperature radiation, specular reflection, and adhesion accumulation on whiteness detection, thus obtaining a pure diffuse reflection image. Based on a multi-task deep learning network, the whiteness value and confidence level of each caustic soda flake are output, and weighted fusion and feedback control are performed according to the confidence level. This achieves online accurate detection and real-time closed-loop control of caustic soda whiteness, solving the problem of positive deviation in whiteness detection on existing caustic soda production lines.

[0007] Preferably, the synchronous acquisition of visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images of caustic soda flakes on the caustic soda flake conveying line, and the subsequent geometric registration, includes the following steps: Visible light multi-angle polarization image, near-infrared multispectral image and thermal infrared image of caustic soda are acquired at a fixed frequency and at the same time above the caustic soda conveying line to form the acquired image; Geometric registration is performed on the acquired images, and pixel coordinate mapping is established based on the calibration board so that all images correspond one-to-one at the pixel level.

[0008] Preferably, the step of retrieving the true temperature distribution on the surface of caustic soda flakes based on the near-infrared multispectral image and the thermal infrared image includes the following steps: A radiation measurement model is established based on the theory of multispectral radiation thermometry. The radiation values ​​of each spectral channel are obtained from the near-infrared multispectral image and the thermal infrared image to form a radiation value matrix. The radiance matrix is ​​decomposed by QR to obtain the initial emissivity value. The matching degree between the measurement vector and the spectrum in the library is calculated by dynamic matching of the emissivity spectrum library. The spectrum with the matching degree reaching the preset value is selected for weighted combination, and the adaptive emissivity is obtained by combining polynomial smoothing constraints. The radiometric values, adaptive emissivity, and polynomial coefficients of each channel in the radiometric matrix are converted into two-dimensional images by Gramian angular field transformation and then stacked into a three-dimensional data cube. The three-dimensional data cube is input into a pre-trained CBAM-ResNet18 convolutional neural network, which outputs the true temperature distribution on the surface of caustic soda flakes.

[0009] Preferably, the step of calculating the thermal radiation-added grayscale value and subtracting it from the visible light multi-angle polarization image to obtain a temperature-corrected image includes the following steps: Based on the actual temperature distribution on the surface of caustic soda flakes, the thermal radiation-added gray value of each pixel is calculated using a pre-calibrated temperature-grayscale increment function. The temperature-corrected image is obtained by subtracting the corresponding thermal radiation-added gray value from the gray value of each pixel in the visible light multi-angle polarization image.

[0010] Preferably, the process of establishing the joint observation equation and solving to separate the diffuse reflectance image includes the following steps: The thermal radiation intensity is determined based on the temperature-corrected image, and the temperature-corrected image is combined with the visible light multi-angle polarization image. The Stokes parameters are calculated based on the visible light multi-angle polarization image to obtain the degree of linear polarization and the polarization angle, and a joint observation equation including the thermal radiation polarization component is established. Independent component analysis is used to initialize the diffuse reflection component and the specular reflection component, and an objective function is constructed. The diffuse reflection component, the specular reflection component, the specular reflection polarization angle and the thermal radiation polarization coefficient are iteratively optimized through an alternating minimization strategy until convergence, and the separated diffuse reflection image is output.

[0011] Preferably, the step of performing image segmentation on the diffuse image, constructing a shape prior based on a statistical shape model, and extracting the mask for each individual caustic soda flake includes the following steps: Offline collection of caustic soda flake sample contours; statistical shape model was constructed through Procrustes analysis and principal component analysis to obtain the average shape and principal components of shape variation. Based on the diffuse reflection image, the position and size of the caustic soda flakes are estimated, and the shape parameters are adjusted to generate a prior level set function. A Chan-Vese level set model with a prior level set function as a shape prior constraint is adopted. An energy functional is defined, and the level set parameters are dynamically adjusted according to the local contrast of the diffuse reflection image, the mean gradient magnitude, and the estimated caustic soda size through fuzzy logic. The level set function is iteratively evolved until convergence, and the watershed algorithm with label control is used to finely segment the adhesion region, outputting the mask for each independent caustic soda flake.

[0012] Preferably, the step of outputting the whiteness value and corresponding confidence level of each caustic soda flake, and calculating the overall whiteness detection result at the current moment by weighted average based on the confidence level, includes the following steps: Visible light multi-angle polarization image, temperature-corrected image, diffuse reflection image and mask are used as multiple inputs. Multi-scale features are extracted by a shared convolutional neural network and fused into a shared feature map. Features are extracted for each region corresponding to the caustic soda mask using RoI Align, and then connected to the whiteness regression branch and the confidence prediction branch respectively, outputting the whiteness value and confidence. The segmentation prediction map is output by the decoder after the shared feature map, and together with the mask, the network training is optimized. The network is trained end-to-end using a joint loss function, including whiteness regression mean squared error, confidence binary cross-entropy, and segmentation Dice loss. During online inference, the whiteness value and confidence level of each caustic soda flake are output, and the overall whiteness detection result at the current moment is calculated by weighted averaging of the confidence levels.

[0013] Preferably, the step of screening the whiteness values ​​and adjusting the sampling frequency based on the confidence level, and adjusting the parameters of the caustic soda production line according to the screened whiteness values, includes the following steps: When the confidence level is higher than the preset threshold, the corresponding whiteness value is included in the statistics; when the confidence level is lower than the preset threshold, the corresponding whiteness value is discarded and an alarm is triggered. When the average confidence level of a consecutive preset number of frames is lower than the preset value, the sampling frequency of the image acquisition module is increased; when the average confidence level recovers to above the preset value, the original frequency is restored. When the confidence level remains below the preset value for more than the preset duration, a device check is performed. The parameters of the caustic soda flake production line, including the drum cooling water volume, scraper gap, and molten salt temperature, are adjusted based on the whiteness value after screening.

[0014] Preferably, locating the abnormal period when a quality anomaly occurs includes the following steps: The overall whiteness test results at each moment, the confidence level of each caustic soda flake, and the corresponding caustic soda flake production line parameters are timestamped and stored. When an abnormality in whiteness is detected, the parameters of the caustic soda flake production line during the same period are retrieved based on the time of the abnormality. By comparing the parameter fluctuations during normal periods, the period of parameter fluctuations in the caustic soda flake production line that caused the abnormality is located.

[0015] This invention provides a closed-loop detection and quality traceability system for whiteness in caustic soda flake production lines. It offers the following advantages: 1. This invention simultaneously acquires multispectral and polarization images and performs multi-level corrections to sequentially eliminate dynamic interference from high-temperature radiation, specular reflection, and adhesion accumulation on whiteness detection, thereby obtaining a pure diffuse reflection image. Based on a multi-task deep learning network, it outputs the whiteness value and confidence level of each caustic soda flake, and performs weighted fusion and feedback control based on the confidence level, thus realizing online accurate detection and real-time closed-loop control of caustic soda whiteness, solving the problem of positive deviation in whiteness detection on existing caustic soda production lines.

[0016] 2. This invention constructs a data foundation for multi-source information fusion by simultaneously acquiring visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images. It uses multispectral thermometry theory to invert the true temperature distribution on the surface of caustic soda flakes, subtracts the additional gray value of thermal radiation from the visible light image, and eliminates the global interference of high-temperature thermal radiation on whiteness measurement. By establishing a joint observation equation that includes the polarization component of thermal radiation, the temperature-corrected image and the polarization image are combined to solve the problem, and a pure diffuse reflection image is separated, effectively suppressing the local specular interference caused by specular reflection.

[0017] 3. This invention constructs a statistical shape model of caustic soda flakes offline to obtain the average shape and principal components of shape change. Based on the diffuse reflection image, it estimates the position and size of the caustic soda flakes to generate a prior level set function. It adopts a Chan-Vese level set model with shape prior constraints and dynamically adjusts the level set parameters according to image features through fuzzy logic, thereby achieving accurate segmentation of adhering caustic soda flakes. Combined with a marker-controlled watershed algorithm, it performs fine segmentation of the adhering region and extracts the mask of each individual caustic soda flake, improving the whiteness detection from the image region level to the individual caustic soda flake level, significantly improving the representativeness and accuracy of the detection.

[0018] 4. This invention constructs a multi-task deep learning network, using visible light multi-angle polarization images, temperature-corrected images, diffuse reflection images, and masks as multiple inputs. After shared feature extraction, it simultaneously outputs the whiteness value and confidence level of each caustic soda flake. It employs three tasks—whiteness regression, confidence prediction, and auxiliary segmentation—for joint optimization, enabling the network to learn more discriminative feature representations. During online inference, the overall whiteness detection result is calculated using a confidence-weighted average, reducing the impact of low-confidence caustic soda flakes on the overall result. It also provides whiteness values ​​and reliability assessments, providing a more robust feedback signal for closed-loop control. Attached Figure Description

[0019] Figure 1 This is an architecture diagram of a whiteness closed-loop detection and quality traceability system for a caustic soda flake production line proposed in this invention; Figure 2 This is a flowchart of a closed-loop detection and quality traceability method for whiteness on a caustic soda flake production line, as proposed in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: In the first embodiment of the present invention, the present invention provides a closed-loop detection and quality traceability system for whiteness of caustic soda flakes production line, such as... Figure 1 As shown, it includes: The image acquisition module is used to simultaneously acquire visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images of caustic soda flakes on the caustic soda flake conveying line, and perform geometric registration. Furthermore, visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images of caustic soda flakes on the caustic soda flake delivery line are acquired simultaneously and geometrically registered, including the following steps: Visible light multi-angle polarization image, near-infrared multispectral image and thermal infrared image of caustic soda are acquired at a fixed frequency and at the same time above the caustic soda conveying line to form the acquired image; Geometric registration is performed on the acquired images, and pixel coordinate mapping is established based on the calibration board so that all images correspond one-to-one at the pixel level.

[0022] Specifically, a multispectral-polarization fusion imaging system, consisting of a visible light camera, a near-infrared camera, a thermal infrared camera, and an electrically driven rotating polarizer holder above the caustic soda flake conveyor line, utilizes an FPGA synchronous controller to trigger all cameras and the polarizer holder at a fixed frequency, simultaneously acquiring visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images of the caustic soda flakes. For visible light multi-angle polarization images, the electrically driven rotating polarizer holder sequentially switches between four polarization directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees, acquiring one image for each direction. For near-infrared multispectral images, the near-infrared camera is equipped with a filter wheel, sequentially switching between filters at three center wavelengths: 850 nm, 1050 nm, and 1250 nm, acquiring near-infrared images in the corresponding bands. The thermal infrared camera operates in the 8 to 14 micrometer band, directly acquiring thermal infrared images.

[0023] Geometric registration was performed on the acquired images, and a checkerboard calibration board was used to pre-establish pixel coordinate mapping relationships. A visible light 0-degree polarized image was selected as the reference image. The pixel coordinates of the checkerboard corner points in each image were extracted through corner detection, and the homography matrix from each image to the reference image was calculated. ,in Let be the pixel coordinates of the image to be registered. , () represents the coordinates after registration. This is the homography matrix. The least squares method is used to solve this matrix, ensuring a one-to-one correspondence between all images at the pixel level, thus providing spatially aligned multi-source image data for subsequent modules.

[0024] The radiation correction module is used to invert the true temperature distribution on the surface of caustic soda flakes based on near-infrared multispectral images and thermal infrared images, calculate the thermal radiation-added gray value and subtract it from the visible light multi-angle polarization image to obtain a temperature-corrected image. Furthermore, the true temperature distribution on the surface of caustic soda flakes is retrieved based on near-infrared multispectral images and thermal infrared images, including the following steps: A radiation measurement model is established based on the theory of multispectral radiation thermometry. The radiation values ​​of each spectral channel are obtained from near-infrared multispectral images and thermal infrared images to form a radiation value matrix. The radiance matrix is ​​decomposed by QR to obtain the initial emissivity value. The matching degree between the measurement vector and the spectrum in the library is calculated by dynamic matching of the emissivity spectrum library. The spectrum with the matching degree reaching the preset value is selected for weighted combination, and the adaptive emissivity is obtained by combining polynomial smoothing constraints. The radiometric values, adaptive emissivity, and polynomial coefficients of each channel in the radiometric matrix are converted into two-dimensional images by Gramian angular field transformation and then stacked into a three-dimensional data cube. The three-dimensional data cube is input into a pre-trained CBAM-ResNet18 convolutional neural network, which outputs the true temperature distribution on the surface of caustic soda flakes.

[0025] Furthermore, the thermal radiation-added grayscale value is calculated and subtracted from the visible light multi-angle polarization image to obtain a temperature-corrected image, including the following steps: Based on the actual temperature distribution on the surface of caustic soda flakes, the thermal radiation-added gray value of each pixel is calculated using a pre-calibrated temperature-grayscale increment function. The temperature-corrected image is obtained by subtracting the corresponding thermal radiation-added gray value from the gray value of each pixel in the visible light multi-angle polarization image.

[0026] Specifically, a radiation measurement model is established based on multispectral radiometric theory. Radiation values ​​for each spectral channel are obtained from near-infrared multispectral images and thermal infrared images, forming a radiation value matrix. Generally, the radiation intensity of an object's surface is a function of its true temperature and spectral emissivity. According to Planck's radiation law, the first... The output radiance of each spectral channel can be expressed as: ,in, Indicates the first Spectral channels in pixel coordinates The radiation value at that location is read directly from a near-infrared multispectral image or a thermal infrared image; This is the channel calibration constant, which is obtained in advance through a blackbody furnace calibration experiment and reflects the spectral response characteristics of the channel; For the target at wavelength and true temperature The spectral emissivity at that wavelength, which ranges from 0 to 1, represents the ratio of the radiative power of the caustic soda surface at that wavelength to the radiative power of a blackbody. The center wavelength is expressed in micrometers. The second radiation constant has a value of 14388 micrometers·Kelvin; For pixels The actual temperature at that location is the target physical quantity to be solved.

[0027] Spectral emissivity Since the emissivity is unknown and varies with wavelength and temperature, the emissivity matrix is ​​first decomposed using QR decomposition to obtain initial values. The emissivity values ​​of each pixel are then used to construct a measurement matrix. ,in This represents the number of spectral channels. This represents the total number of pixels. QR decomposition is performed on the measurement matrix. ,in It is an orthogonal matrix. Given an upper triangular matrix, solve for... The matrix obtains the feature emissivity vector for each pixel, which serves as an initial emissivity estimate.

[0028] To further improve the accuracy of emissivity estimation, a fine-grained correction of emissivity is performed through dynamic matching with an emissivity spectral library. An emissivity spectral library is pre-established. Each These spectra correspond to the emissivity spectra under typical operating conditions, obtained through offline measurements of caustic soda flakes of different temperatures and purities. For the measurement vector of the current pixel... Calculate its matching degree with each spectrum in the library: The three spectra with the highest matching degree are selected and weighted together, with the weights obtained by normalizing the matching degree to form an adaptive emissivity. Simultaneously, multinomial regression is introduced to smooth the emissivity. ,in The fitting coefficients are determined by the least squares method. This polynomial constraint ensures that the emissivity changes with wavelength in accordance with physical laws.

[0029] The radiance values ​​of each channel, adaptive emissivity, and polynomial coefficients in the radiance matrix are converted into two-dimensional images using Gramian angular field transform, and then stacked into a three-dimensional data cube. Gramian angular field transform is a method for converting one-dimensional sequences into two-dimensional images, which can preserve the temporal dependencies and structural information of the data. For one-dimensional sequences... First, normalization is performed: Then calculate the angle corresponding to each point. The elements of the Gramian angular field matrix G are defined as follows: This matrix reflects the angular relationship between any two points in the sequence. Gramian angular field transformation is applied to the multi-channel radiance, adaptive emissivity, and polynomial coefficients of each pixel to obtain three 64×64 two-dimensional images, which are then stacked to form a three-dimensional data cube. .

[0030] A 3D data cube is input into a pre-trained CBAM-ResNet18 convolutional neural network, which outputs the true temperature distribution on the surface of caustic soda flakes. CBAM-ResNet18 is an improved network based on ResNet18, introducing a convolutional block attention module, which can simultaneously focus on features in both channel and spatial dimensions. The network takes a 3D data cube as input, extracts features through multiple residual blocks and attention modules, and finally outputs the temperature value of each pixel through a fully connected layer. During network training, the true temperature calibrated by a blackbody furnace is used as the label, and the loss function is mean squared error. After training, the network can adaptively reproduce an accurate temperature distribution map based on the multidimensional features of the input. .

[0031] Based on the retrieved true temperature distribution of the caustic soda surface, the thermal radiation-added gray value of each pixel was calculated using a pre-calibrated temperature-grayscale increment function. Calibration experiments were conducted using a standard white board at different temperatures, and the increment of the grayscale value of the visible light camera as a function of temperature was measured. A linear function was then fitted to obtain the result. ,in This is the proportionality coefficient. Reference temperature. Additional grayscale value for thermal radiation. This reflects the interference component of high-temperature thermal radiation on visible light images.

[0032] The temperature-corrected image is obtained by subtracting the corresponding thermal radiation-added gray value from the gray value of each pixel in the visible light multi-angle polarization image. This correction operation eliminates the global brightness shift caused by high-temperature thermal radiation, ensuring that the temperature-corrected image contains only the reflection information from the caustic soda surface, providing an accurate base image for subsequent polarization-thermal radiation coupling correction. For multi-angle polarized images, the same radiometric correction process is performed on the image at each angle to ensure temperature consistency across all polarization directions.

[0033] The coupling correction module is used to combine the temperature-corrected image with the visible light multi-angle polarization image, establish a joint observation equation, and solve to separate the diffuse reflection image; Furthermore, the joint observation equation is established and solved to separate the diffuse reflectance image, including the following steps: The thermal radiation intensity is determined based on the temperature-corrected image, and the temperature-corrected image is combined with the visible light multi-angle polarization image. The Stokes parameters are calculated based on the visible light multi-angle polarization image to obtain the degree of linear polarization and the polarization angle. A joint observation equation including the thermal radiation polarization component is then established. Independent component analysis is used to initialize the diffuse reflection component and the specular reflection component, and an objective function is constructed. The diffuse reflection component, the specular reflection component, the specular reflection polarization angle and the thermal radiation polarization coefficient are iteratively optimized through an alternating minimization strategy until convergence, and the separated diffuse reflection image is output.

[0034] Specifically, the thermal radiation intensity is determined based on the temperature-corrected image, and then combined with a visible light multi-angle polarization image. Stokes parameters are calculated from the visible light multi-angle polarization image to obtain the degree of linear polarization and the polarization angle, and a joint observation equation incorporating the polarization component of thermal radiation is established. Specifically, the visible light multi-angle polarization image includes images with four polarization directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees, denoted as... According to polarization optics theory, the Stokes vector is used to describe the polarization state of light, and the Stokes parameters are defined as follows: ; ; ; in, The total light intensity represents the sum of the light intensities in all polarization directions at that pixel location. and The linear polarization components are described, representing the polarization differences in the horizontal and vertical directions and the polarization difference at 45 degrees. The circular polarization component is usually negligible in the reflection from the caustic soda surface. The degree of linear polarization and polarization angle are further calculated based on the Stokes parameters: ; ; linear polarization degree This represents the proportion of linearly polarized components in a light wave, ranging from 0 to 1, where 0 represents completely unpolarized light and 1 represents completely linearly polarized light. Polarization angle. This represents the angle between the vibration direction of linearly polarized light and the horizontal direction. These two parameters reflect the polarization characteristics of light reflected from the surface of caustic soda flakes. Specular reflection light typically exhibits a high degree of linear polarization, while diffuse reflection light exhibits a low degree of linear polarization or non-polarized characteristics.

[0035] A joint observation equation incorporating the polarization component of thermal radiation is established. Considering that the thermal radiation from high-temperature caustic soda flakes also possesses polarization characteristics, and that these characteristics vary with the observation angle, the polarization component of thermal radiation is introduced into the observation equation. The joint observation equation is expressed as: ,in, Polarization angle The grayscale value of the visible light image is provided by the image acquisition module. The diffuse reflection component represents the intensity of diffuse reflected light from the surface of caustic soda flakes. This component is independent of the polarization angle and is the basic information required for subsequent whiteness detection. The specular reflection component represents the intensity of specularly reflected light from the surface of caustic soda flakes, and this component has complete polarization characteristics. The polarization angle of the specular reflection indicates the polarization direction of the reflected light. Temperature distribution obtained from the radiation correction module represents thermal radiation intensity. It is obtained by calibration curve conversion, reflecting the intensity of high-temperature thermal radiation in the visible light band. The thermal radiation polarization coefficient describes the polarization modulation characteristics of thermal radiation. It is related to the surface material properties of caustic soda flakes and is one of the parameters to be solved in this equation.

[0036] This joint observation equation decomposes the reflected light from the caustic soda surface into three parts: a diffuse reflection component independent of angle, a specular reflection component varying according to a cosine square law, and a thermal radiation polarization component varying according to a double-angle cosine law. These unknown parameters can be solved using observations at multiple polarization angles.

[0037] Independent component analysis (ICA) was used to initialize the diffuse and specular reflection components. ICA is a blind source separation method that assumes the observed signal is a linear mixture of multiple independent source signals. The images from four polarization angles were used to construct the observation matrix. ,in Let be the total number of pixels. The independent component analysis model can be expressed as: ,in, It is a mixed matrix. Let be the source signal matrix, containing diffuse reflection and specular reflection components. The model is solved using the FastICA algorithm to obtain initial estimates for the diffuse and specular reflection components, as well as the initial value of the specular reflection polarization angle.

[0038] Construct an objective function and iteratively optimize the diffuse reflection component, specular reflection component, specular reflection polarization angle, and thermal radiation polarization coefficient using an alternating minimization strategy. The objective function is defined as follows: ,in, This represents the summation over the four polarization angles. and These are the diffuse reflection component, specular reflection component, specular reflection polarization angle, and thermal radiation polarization coefficient, respectively, which are to be optimized. The thermal radiation intensity calculated based on the temperature distribution remains constant during the optimization process. This is the regularization coefficient, which controls the weight of the temperature smoothing term. The square of the temperature gradient modulus represents the smoothness of the temperature field. This regularization term can constrain the spatial variation of the thermal radiation polarization coefficient and avoid overfitting.

[0039] The specific implementation steps of the alternating minimization strategy are as follows: Step 1, fix and ,renew At this point, the objective function simplifies to... The linear least squares problem can be solved directly. The updated value. The second step is to fix it. ,renew and At this point, the objective function is non-linear, and can be optimized using gradient descent or the Levenberg-Marquardt algorithm. Repeat the above two steps until the change in the objective function is less than a preset threshold or the maximum number of iterations is reached.

[0040] To improve the stability and convergence speed of the solution, constraints are imposed on the initial values. For example, the specular reflection component. It should be a non-negative value; non-negativity constraints can be applied during the optimization process. Thermal radiation polarization coefficient. The value range is generally between -0.5 and 0.5, and the upper and lower limits can be preset according to the material properties.

[0041] After iterative convergence, the separated diffuse reflection image is output. The image has eliminated interference from specular reflection and thermal radiation polarization components, retaining only diffuse reflectance information that reflects the intrinsic color of caustic soda flakes. The diffuse reflectance image is the fundamental input for subsequent adhesion segmentation and whiteness detection modules, and its quality directly affects the accuracy of the final whiteness detection.

[0042] The quality of the separated diffuse reflection image is verified by calculating its residual with the original visible light image. If the residual is too large, it indicates that the separation effect is not good. The regularization coefficient can be adjusted or the image can be re-initialized. This achieves the joint decoupling of high-temperature thermal radiation and specular reflection, providing a clean base image for the accurate detection of caustic soda whiteness.

[0043] The adhesion segmentation module is used to segment diffuse reflection images, construct shape priors based on statistical shape models, and extract the mask for each individual caustic soda flake. Furthermore, image segmentation is performed on the diffuse reflection image, a shape prior is constructed based on a statistical shape model, and a mask for each individual caustic soda flake is extracted, including the following steps: Offline collection of caustic soda flake sample contours; statistical shape model was constructed through Procrustes analysis and principal component analysis to obtain the average shape and principal components of shape variation. Based on the diffuse reflection image, the location and size of the caustic soda flakes are estimated, and the shape parameters are adjusted to generate a prior level set function; A Chan-Vese level set model with a prior level set function as a shape prior constraint is adopted. An energy functional is defined, and the level set parameters are dynamically adjusted according to the local contrast of the diffuse reflection image, the mean gradient magnitude, and the estimated caustic soda size through fuzzy logic. The level set function is iteratively evolved until convergence, and the watershed algorithm with label control is used to finely segment the adhesion region, outputting the mask for each independent caustic soda flake.

[0044] Specifically, in this embodiment, the contours of caustic soda flake samples are first collected offline, and a statistical shape model is constructed using Procrustes analysis and principal component analysis. Each contour is represented by p uniformly sampled points, denoted as p. ,in For sample index, Set the value to 100. After Procrustes alignment of all contours, calculate the average shape. , Given the total number of samples, construct the covariance matrix. Perform principal component analysis ,in The eigenvector matrix, Let be the eigenvalue diagonal matrix. Retaining the first t principal components, where t is 5, the profile of any caustic soda flake can be represented as... ,in For shape transformation matrix, For shape parameters, follow .

[0045] Based on the diffuse reflectance image, the location and size of the caustic soda flakes are estimated. The average shape is aligned to the candidate region through a similarity transformation, and a prior level set function is generated through a distance transformation. Distance transformation is defined as pixel point The shortest Euclidean distance to the prior contour, with negative values ​​inside the contour and positive values ​​outside the contour.

[0046] Using the Chan-Vese level set model with shape prior constraints, the energy functional is defined as: ; in For level set functions, This refers to the image domain. For regularization Heaviside functions, This is the Dirac delta function for regularization. and These are the average gray values ​​inside and outside the curve, respectively, based on the current... Updated in real time. The length term is used to control the smoothness of the contour. To control the expansion trend by weighting the area term, and The weight for grayscale consistency is usually set to 1. The shape prior weights.

[0047] The system dynamically adjusts based on local contrast, average gradient magnitude, and estimated caustic soda size using fuzzy logic. and The evolution equation for the level set function is: The finite difference method is used for iterative solution until convergence. For severely adhered regions, a marker-controlled watershed algorithm is used for fine segmentation, outputting a mask for each individual caustic soda flake. This represents the total number of caustic soda flakes in the current image.

[0048] The detection module is used to input visible light multi-angle polarization image, temperature-corrected image, diffuse reflection image and mask into a pre-trained multi-task deep learning network, output the whiteness value and corresponding confidence score of each caustic soda flake, and calculate the overall whiteness detection result at the current time based on the weighted average of the confidence scores. Furthermore, the whiteness value and corresponding confidence level of each caustic soda flake are output, and the overall whiteness detection result at the current moment is calculated by weighted average based on the confidence level, including the following steps: Visible light multi-angle polarization image, temperature-corrected image, diffuse reflection image and mask are used as multiple inputs. Multi-scale features are extracted by a shared convolutional neural network and fused into a shared feature map. Features are extracted for each region corresponding to the caustic soda mask using RoI Align, and then connected to the whiteness regression branch and the confidence prediction branch respectively, outputting the whiteness value and confidence. The segmentation prediction map is output by the decoder after the shared feature map, and together with the mask, the network training is optimized. The network is trained end-to-end using a joint loss function, including whiteness regression mean squared error, confidence binary cross-entropy, and segmentation Dice loss. During online inference, the whiteness value and confidence level of each caustic soda flake are output, and the overall whiteness detection result at the current moment is calculated by weighted averaging of the confidence levels.

[0049] Specifically, in this embodiment, visible light multi-angle polarization images are stacked into a multi-channel input. Temperature-corrected image diffuse reflection image and mask As multiple inputs, multi-scale features are extracted through a shared ResNet50 backbone network, and channel concatenation and fusion are performed at each scale to obtain a shared feature map. .

[0050] For each caustic soda mask The corresponding regions are used to extract fixed-size features through RoI Align, and then the feature vectors are obtained through a fully connected layer. Input this vector into the whiteness regression branch and the confidence prediction branch respectively, and output the whiteness value. and confidence level .

[0051] The decoder outputs a segmentation prediction map following the shared feature map. The training is optimized together with the mask. The joint loss function is defined as: Among them, whiteness regression loss , For the first The true whiteness value of each caustic soda flake was obtained by measurement using a laboratory whiteness meter. (Confidence loss) , For confidence level labels, if but otherwise , The preset threshold is used for segmentation loss. , and This is the balance coefficient.

[0052] During online inference, output the whiteness value of each caustic soda flake. and confidence level The overall whiteness test results were calculated using a confidence-weighted average: This result serves as the whiteness detection value at the current moment, and is used for subsequent feedback control and quality traceability.

[0053] The feedback module is used to filter whiteness values ​​and adjust the sampling frequency based on confidence level, and adjust the parameters of the caustic soda production line according to the filtered whiteness values ​​to achieve closed-loop control of whiteness. Furthermore, the whiteness values ​​are screened and the sampling frequency is adjusted based on the confidence level. The parameters of the caustic soda flake production line are then adjusted according to the screened whiteness values, including the following steps: When the confidence level is higher than the preset threshold, the corresponding whiteness value is included in the statistics; when the confidence level is lower than the preset threshold, the corresponding whiteness value is discarded and an alarm is triggered. When the average confidence level of a consecutive preset number of frames is lower than the preset value, the sampling frequency of the image acquisition module is increased; when the average confidence level recovers to above the preset value, the original frequency is restored. When the confidence level remains below the preset value for more than the preset duration, a device check is performed. The parameters of the caustic soda flake production line, including the drum cooling water volume, scraper gap, and molten salt temperature, are adjusted based on the whiteness value after screening.

[0054] Specifically, the feedback module first filters the whiteness value of each caustic soda flake based on confidence level. A high confidence threshold is set. and low confidence threshold ,in Take 0.95, Take 0.80. For each caustic soda whiteness value output by the detection module. and its confidence level ,like If so, the whiteness value is accepted and included in the statistics; If the whiteness value is not found, the whiteness value will be discarded and an alarm signal will be triggered.

[0055] The sampling frequency of the image acquisition module is dynamically adjusted based on the average confidence score across multiple consecutive frames. The average confidence score of the previous N frames at the current time is calculated. ,in For the first The average confidence level of all caustic soda flakes in the frame. When When the frequency is lower than the first preset value, the sampling frequency will be reduced from the reference frequency. Increase to ;when When the confidence level recovers to or above the second preset value, the original sampling frequency is restored. If the confidence level remains below the third preset value for more than a preset duration, the backup detection algorithm is activated and maintenance personnel are notified to inspect the equipment.

[0056] The parameters of the caustic soda flake production line are adjusted based on the screened whiteness values ​​to achieve closed-loop control of whiteness. The screened whiteness values ​​represent the overall whiteness test results at the current moment. The whiteness is obtained by the detection module through a confidence-weighted average. Using caustic soda whiteness as the target control variable, a PID controller is used to adjust the front-end process parameters. Whiteness deviation is defined. ,in For the target whiteness value, For time. The output control quantity of the PID controller. Represented as: ,in This is a proportionality coefficient that determines the strength of the response to the current deviation; These are the integral coefficients used to eliminate steady-state errors; This is the differential coefficient, used to suppress the trend of deviation change. Control quantity. These correspond to the adjustment increments of the drum cooling water valve opening, scraper gap adjustment, and molten salt temperature setpoint, respectively. Through multiple parallel PID controllers, the whiteness of the caustic soda flakes is stabilized near the target value, achieving online quality optimization of the production process.

[0057] The quality traceability module is used to timestamp and store the overall whiteness test results, confidence levels, and corresponding caustic soda production line parameters at each moment, and to locate the abnormal period when quality is abnormal.

[0058] Furthermore, locating the abnormal period when a quality anomaly occurs includes the following steps: The overall whiteness test results at each moment, the confidence level of each caustic soda flake, and the corresponding caustic soda flake production line parameters are timestamped and stored. When an abnormality in whiteness is detected, the parameters of the caustic soda flake production line during the same period are retrieved based on the time of the abnormality. By comparing the parameter fluctuations during normal periods, the period of parameter fluctuations in the caustic soda flake production line that caused the abnormality is located.

[0059] Specifically, the quality traceability module will record the overall whiteness detection results at each moment. Confidence levels of each alkali tablet And the corresponding parameters of the caustic soda flake production line, including drum cooling water volume, scraper gap, molten salt temperature, etc., with a unified timestamp. The index is stored in relation to other data. All data is stored in the quality traceability database in chronological order.

[0060] When an abnormal whiteness is detected, parameters from the concurrent caustic soda flake production line are retrieved based on the time of the abnormality. The criteria for determining an abnormal whiteness are as follows: ,in and This defines the lower and upper limits of the acceptable whiteness range. It records the start and end times of any anomalies and extracts the time-series data of parameters for that time period and its preceding and following extended periods from the database.

[0061] For each parameter, calculate its mean and standard deviation over a normal historical period, where the normal historical period is selected as data under the same operating conditions before the anomaly occurred. Then, calculate the parameter sequence over the anomaly period. Calculate the standardized deviation .when When parameters continuously exceed a preset threshold, that period is marked as a parameter fluctuation period. By analyzing the temporal correlation between the fluctuations of each parameter and the whiteness anomaly, the period of parameter fluctuation in the caustic soda flake production line causing the anomaly can be located.

[0062] Example 2: In a second embodiment of the present invention, the present invention provides a closed-loop detection and quality traceability method for whiteness on a caustic soda flake production line, such as... Figure 2 As shown, it includes the following steps: Simultaneously acquire visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images of caustic soda flakes on the caustic soda flake delivery line, and perform geometric registration; The true temperature distribution on the surface of caustic soda flakes is retrieved from near-infrared multispectral images and thermal infrared images. The thermal radiation-added gray value is calculated and subtracted from the visible light multi-angle polarization image to obtain a temperature-corrected image. By combining temperature-corrected images with visible light multi-angle polarization images, a joint observation equation is established and solved to separate the diffuse reflection image. The diffuse reflection image is segmented, a shape prior is constructed based on a statistical shape model, and the mask for each individual caustic soda flake is extracted. The pre-trained multi-task deep learning network inputs visible light multi-angle polarization images, temperature-corrected images, diffuse reflection images, and masks to output the whiteness value and corresponding confidence level of each caustic soda flake, and calculates the overall whiteness detection result at the current time by weighted average based on the confidence level. The whiteness values ​​are screened and the sampling frequency is adjusted based on the confidence level. The parameters of the caustic soda production line are adjusted according to the screened whiteness values ​​to achieve closed-loop control of whiteness. The overall whiteness test results, confidence levels, and corresponding caustic soda production line parameters at each moment are timestamped and stored to locate the abnormal period when quality is abnormal.

[0063] In a caustic soda flake production workshop with an annual output of 100,000 tons, the production line operates continuously for extended periods. High-temperature radiation, surface reflection, and adhesion between caustic soda flakes during the initial forming stage cause drastic fluctuations in online whiteness detection data, while offline testing results are severely delayed, failing to provide timely guidance for process adjustments and resulting in a large number of defective products. To address these issues, this invention provides a closed-loop whiteness detection and quality traceability method for caustic soda flake production lines. The process is as follows: Figure 2 As shown. The specific implementation process of this method is as follows: First, visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images of caustic soda flakes were acquired simultaneously and geometrically registered. Based on the multispectral radiometric theory, the true temperature distribution of caustic soda surface is retrieved from near-infrared and thermal infrared images, and the thermal radiation-added gray value in the visible light image is subtracted to obtain a temperature-corrected image. By combining temperature-corrected images with polarization images, a joint observation equation containing the polarization component of thermal radiation is established, and the diffuse reflection image is separated by alternately minimizing the solution. A shape prior is constructed using a statistical shape model, and an adaptive level set method is used to segment the diffuse reflection image to extract the mask for each independent caustic soda flake. The visible light polarization image, temperature-corrected image, diffuse reflection image, and mask are input into a multi-task deep learning network, which simultaneously outputs the whiteness value and confidence score of each caustic soda flake. The overall whiteness detection result at the current moment is obtained by weighted averaging of the confidence scores. Whiteness values ​​are filtered based on confidence level, low-confidence data are discarded and an alarm is triggered, and the sampling frequency is dynamically adjusted. The filtered whiteness values ​​are used as feedback signals to adjust the drum cooling water volume, scraper gap and molten salt temperature to achieve closed-loop control of whiteness. Finally, all test results and process parameter timestamps are aligned and stored in the database. When whiteness is abnormal, the cause of the abnormality can be accurately located by searching for the fluctuation period of the process parameters in the same period.

[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A closed-loop detection and quality traceability system for whiteness on a caustic soda flake production line, characterized in that, include: The image acquisition module is used to simultaneously acquire visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images of caustic soda flakes on the caustic soda flake conveying line, and perform geometric registration. The radiation correction module is used to invert the true temperature distribution of the caustic soda surface based on the near-infrared multispectral image and the thermal infrared image, calculate the thermal radiation-added gray value and subtract it from the visible light multi-angle polarization image to obtain a temperature-corrected image. The coupling correction module is used to combine the temperature-corrected image with the visible light multi-angle polarization image, establish a joint observation equation, and solve to separate the diffuse reflection image. The adhesion segmentation module is used to segment the diffuse reflection image, construct shape priors based on a statistical shape model, and extract the mask for each individual caustic soda flake. The detection module is used to input the visible light multi-angle polarization image, temperature-corrected image, diffuse reflection image and the mask into a pre-trained multi-task deep learning network, output the whiteness value and corresponding confidence level of each caustic soda flake, and calculate the overall whiteness detection result at the current time based on the weighted average of the confidence level. The feedback module is used to filter the whiteness value and adjust the sampling frequency according to the confidence level, and adjust the parameters of the caustic soda production line according to the filtered whiteness value to perform closed-loop control of whiteness. The quality traceability module is used to timestamp and store the overall whiteness test results, confidence levels, and corresponding caustic soda production line parameters at each moment, and to locate the abnormal period when quality is abnormal.

2. The whiteness closed-loop detection and quality traceability system for a caustic soda flake production line according to claim 1, characterized in that: The synchronous acquisition of visible light multi-angle polarization images, near-infrared multispectral images, and thermal infrared images of caustic soda flakes on the caustic soda flake delivery line, and the subsequent geometric registration, includes the following steps: Visible light multi-angle polarization image, near-infrared multispectral image and thermal infrared image of caustic soda are acquired at a fixed frequency and at the same time above the caustic soda conveying line to form the acquired image; Geometric registration is performed on the acquired images, and pixel coordinate mapping is established based on the calibration board so that all images correspond one-to-one at the pixel level.

3. The whiteness closed-loop detection and quality traceability system for a caustic soda flake production line according to claim 1, characterized in that: The process of retrieving the true temperature distribution on the surface of caustic soda flakes based on the near-infrared multispectral image and thermal infrared image includes the following steps: A radiation measurement model is established based on the theory of multispectral radiation thermometry. The radiation values ​​of each spectral channel are obtained from the near-infrared multispectral image and the thermal infrared image to form a radiation value matrix. The radiance matrix is ​​decomposed by QR to obtain the initial emissivity value. The matching degree between the measurement vector and the spectrum in the library is calculated by dynamic matching of the emissivity spectrum library. The spectrum with the matching degree reaching the preset value is selected for weighted combination, and the adaptive emissivity is obtained by combining polynomial smoothing constraints. The radiometric values, adaptive emissivity, and polynomial coefficients of each channel in the radiometric matrix are converted into two-dimensional images by Gramian angular field transformation and then stacked into a three-dimensional data cube. The three-dimensional data cube is input into a pre-trained CBAM-ResNet18 convolutional neural network, which outputs the true temperature distribution on the surface of caustic soda flakes.

4. The whiteness closed-loop detection and quality traceability system for a caustic soda flake production line according to claim 1, characterized in that: The process of calculating the thermal radiation-added grayscale value and subtracting it from the visible light multi-angle polarization image to obtain a temperature-corrected image includes the following steps: Based on the actual temperature distribution on the surface of caustic soda flakes, the thermal radiation-added gray value of each pixel is calculated using a pre-calibrated temperature-grayscale increment function. The temperature-corrected image is obtained by subtracting the corresponding thermal radiation-added gray value from the gray value of each pixel in the visible light multi-angle polarization image.

5. The whiteness closed-loop detection and quality traceability system for a caustic soda flake production line according to claim 1, characterized in that: The process of establishing the joint observation equation and solving it to separate the diffuse reflectance image includes the following steps: The thermal radiation intensity is determined based on the temperature-corrected image, and the temperature-corrected image is combined with the visible light multi-angle polarization image. The Stokes parameters are calculated based on the visible light multi-angle polarization image to obtain the degree of linear polarization and the polarization angle, and a joint observation equation including the thermal radiation polarization component is established. Independent component analysis is used to initialize the diffuse reflection component and the specular reflection component, and an objective function is constructed. The diffuse reflection component, the specular reflection component, the specular reflection polarization angle and the thermal radiation polarization coefficient are iteratively optimized through an alternating minimization strategy until convergence, and the separated diffuse reflection image is output.

6. The whiteness closed-loop detection and quality traceability system for a caustic soda flake production line according to claim 1, characterized in that: The process of segmenting the diffuse image, constructing a shape prior based on a statistical shape model, and extracting the mask for each individual caustic soda flake includes the following steps: Offline collection of caustic soda flake sample contours; statistical shape model was constructed through Procrustes analysis and principal component analysis to obtain the average shape and principal components of shape variation. Based on the diffuse reflection image, the position and size of the caustic soda flakes are estimated, and the shape parameters are adjusted to generate a priori level set function; A Chan-Vese level set model with a prior level set function as a shape prior constraint is adopted. An energy functional is defined, and the level set parameters are dynamically adjusted according to the local contrast of the diffuse image, the mean gradient magnitude, and the estimated caustic soda size through fuzzy logic. The level set function is iteratively evolved until convergence, and the watershed algorithm with label control is used to finely segment the adhesion region, outputting the mask for each independent caustic soda flake.

7. The whiteness closed-loop detection and quality traceability system for a caustic soda flake production line according to claim 1, characterized in that: The process of outputting the whiteness value and corresponding confidence level of each caustic soda flake, and calculating the overall whiteness detection result at the current moment by weighted average based on the confidence level, includes the following steps: Visible light multi-angle polarization image, temperature-corrected image, diffuse reflection image and mask are used as multiple inputs. Multi-scale features are extracted by a shared convolutional neural network and fused into a shared feature map. Features are extracted for each region corresponding to the caustic soda mask using RoI Align, and then connected to the whiteness regression branch and the confidence prediction branch respectively, outputting the whiteness value and confidence. The segmentation prediction map is output by the decoder after the shared feature map, and together with the mask, the network training is optimized. The network is trained end-to-end using a joint loss function, including whiteness regression mean squared error, confidence binary cross-entropy, and segmentation Dice loss. During online inference, the whiteness value and confidence level of each caustic soda flake are output, and the overall whiteness detection result at the current moment is calculated by weighted averaging of the confidence levels.

8. The whiteness closed-loop detection and quality traceability system for a caustic soda flake production line according to claim 1, characterized in that: The process of filtering the whiteness values ​​based on the confidence level and adjusting the sampling frequency, and then adjusting the parameters of the caustic soda production line based on the filtered whiteness values, includes the following steps: When the confidence level is higher than the preset threshold, the corresponding whiteness value is included in the statistics; when the confidence level is lower than the preset threshold, the corresponding whiteness value is discarded and an alarm is triggered. When the average confidence level of a consecutive preset number of frames is lower than the preset value, the sampling frequency of the image acquisition module is increased; when the average confidence level recovers to above the preset value, the original frequency is restored. When the confidence level remains below the preset value for more than the preset duration, a device check is performed. The parameters of the caustic soda flake production line, including the drum cooling water volume, scraper gap, and molten salt temperature, are adjusted based on the whiteness value after screening.

9. The whiteness closed-loop detection and quality traceability system for a caustic soda flake production line according to claim 1, characterized in that: Locating the abnormal period when a quality anomaly occurs includes the following steps: The overall whiteness test results at each moment, the confidence level of each caustic soda flake, and the corresponding caustic soda flake production line parameters are timestamped and stored. When an abnormality in whiteness is detected, the parameters of the caustic soda flake production line during the same period are retrieved based on the time of the abnormality. By comparing the parameter fluctuations during normal periods, the period of parameter fluctuations in the caustic soda flake production line that caused the abnormality is located.

Images