A low porosity anode carbon block porosity real-time estimation system and low porosity anode carbon block finished product thereof

Abstract

The application provides a low-porosity anode carbon block porosity real-time estimation system and a low-porosity anode carbon block finished product, relates to the material nondestructive testing technical field, and through the model uncertainty generated by the preliminary optical global scanning, intelligently guides the calibration point position selection of the high-precision physical measurement, and utilizes the high-information-value calibration data to perform real-time, incremental online optimization on the mapping model. Not only can the "model drift" problem commonly existing in industrial production be effectively overcome, and continuous high estimation accuracy be ensured; but also through an active calibration strategy, the maximum model correction value is obtained at the lowest calibration cost. Finally, a multi-dimensional estimation confidence evaluation is introduced, which provides solid technical support for realizing truly reliable and reliable industrial intelligent quality control.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing technology for materials, specifically to a real-time porosity estimation system for low-porosity anode carbon blocks and a finished low-porosity anode carbon block product. Background Technology

[0002] In the high-end materials manufacturing and process industries, a profound transformation is underway, shifting from traditional offline quality inspection to online, real-time, data-driven process analysis (PAT) technology. Building digital twin systems capable of accurately reflecting key internal quality indicators of products in real time has become a crucial technological direction for improving production efficiency, ensuring product consistency, and enhancing core competitiveness. Particularly in the production of functional materials with stringent performance requirements, non-destructive, comprehensive, and high-precision online characterization of core physical properties is the technological cornerstone for achieving intelligent manufacturing.

[0003] Despite some progress in this field, the following core challenges remain in achieving real-time and accurate estimation of key internal physical properties such as the porosity of anode carbon blocks:

[0004] Existing technologies, such as the scheme in Publication No. CN109239074B, demonstrate the approach of using machine vision for surface geometric feature detection. These methods generally rely on offline-trained static inference models. However, in actual production, batch changes in raw materials and minute fluctuations in process parameters can cause a slow but continuous "drift" in the mapping relationship between the product's optical features and internal physical properties. Static models cannot adaptively adjust online, resulting in inherent limitations in their long-term accuracy and reliability.

[0005] Calibrling optical models typically requires precise but time-consuming and destructive physical measurements. Existing technologies face bottlenecks in effectively utilizing this valuable physical calibration information. Model updates are often performed offline and in batches, or fixed, indiscriminate sampling strategies are employed, leading to slow responses to dynamic changes during production. For example, while the technical solution in publication number CN118313843A addresses the traceability of carbon blocks, it does not address how to intelligently decide where to perform physical testing based on the real-time status of the product to obtain information most efficiently. This makes it difficult to achieve an optimal balance between calibration costs and information value. Summary of the Invention

[0006] The purpose of this invention is to provide a real-time porosity estimation system for low-porosity anode carbon blocks and a finished low-porosity anode carbon block, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A real-time porosity estimation system for low-porosity anode carbon blocks, comprising the following steps:

[0009] The optical data acquisition module is used to acquire optical microstructure data of the surface of the anode carbon block under test;

[0010] The physical property acquisition module is used to acquire gas transport characteristic data at at least one calibration point of the anode carbon block;

[0011] The calibration decision module is used to generate an initial porosity distribution map based on the optical micromorphology data using an online-updable mapping model, and to calculate and output an optimal calibration point instruction according to the preset information value criteria of the initial porosity distribution map.

[0012] The model optimization module is used to acquire the gas transport characteristic data at the calibration point determined by the optimal calibration point instruction, and to incrementally optimize the online-updable mapping model using the gas transport characteristic data; and,

[0013] The result generation module is used to process the optical micromorphology data using the incrementally optimized mapping model to generate a corrected porosity distribution map and an estimated confidence distribution map.

[0014] Compared with existing technologies, the beneficial effects of this invention are: by constructing an integrated closed-loop adaptive estimation framework of "perception-decision-learning". This framework no longer treats optical detection and physical detection as two independent links, but rather enables them to work collaboratively and mutually benefit through an innovative information flow closed loop. Specifically, it addresses the limitations of existing technologies in a targeted manner:

[0015] By introducing an online incremental optimization mechanism, the core mapping model is continuously and online fine-tuned during production using newly acquired physical calibration data and algorithms such as transfer learning or Bayesian updates. This enables the model to track and compensate for "model drift" caused by process fluctuations in real time, thereby maintaining high estimation accuracy over a long period in dynamic production environments.

[0016] A proactive calibration decision-making strategy based on maximizing information value is proposed. The current model is used to perform a rapid preliminary assessment of the entire carbon block, generating a "model uncertainty map" or "feature novelty map." This map reveals in a graphical way which regions the model predicts the worst features. Then, a decision is made to perform precise physical measurements at these points with the highest information value. This proactive calibration method ensures that every time-consuming and labor-intensive physical measurement provides the model with maximum information gain, improving calibration efficiency. By organically combining proactive calibration decision-making with online incremental optimization—two core innovations—it achieves an evolution from static, unidirectional prediction to dynamic, bidirectional intelligent estimation with self-awareness and self-evolutionary capabilities. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall system module flow of the present invention.

[0018] Figure 2 This is a schematic diagram of the execution logic of the optical data acquisition module, physical characteristic acquisition module, and calibration decision module of the present invention;

[0019] Figure 3 This is a schematic diagram of the structured light coding pattern, the original distorted image, the absolute phase map, and the three-dimensional point cloud data of the present invention.

[0020] Figure 4 This is a schematic diagram illustrating the execution logic of the model optimization module and the result generation module of this invention. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Example 1: Please refer to Figures 1 to 4 The present invention provides a technical solution:

[0024] A real-time porosity estimation system for low-porosity anode carbon blocks includes:

[0025] The optical data acquisition module is used to acquire optical microstructure data of the surface of the anode carbon block under test;

[0026] The physical property acquisition module is used to acquire gas transport characteristic data at at least one calibration point of the anode carbon block;

[0027] The calibration decision module is used to generate an initial porosity distribution map based on optical micromorphology data using an online-updable mapping model, and calculate and output an optimal calibration point instruction according to the preset information value criteria of the initial porosity distribution map.

[0028] The model optimization module is used to acquire gas transport characteristic data at the calibration points determined by the optimal calibration point instruction, and to incrementally optimize the online-updable mapping model using this gas transport characteristic data; and,

[0029] The results generation module processes optical micromorphology data using an incrementally optimized mapping model to generate a corrected porosity distribution map and an estimated confidence distribution map.

[0030] It should be noted that: Figure 3 Figure P1 illustrates a structured light coded pattern used to acquire the three-dimensional morphology of the anode carbon block surface. Specifically, the pattern is a sinusoidal fringe pattern, characterized by a series of parallel vertical stripes. Unlike binarized stripes, the light intensity of the stripes in the figure follows a sinusoidal function distribution in space, represented by a smooth grayscale gradient from pure black to pure white. This coded pattern is projected onto the surface of the object under test, and by analyzing the deformation caused by the surface's height variations, the three-dimensional coordinate information of the surface is calculated. It serves as the initial information carrier for the entire optical data acquisition module.

[0031] Figure 3 Figure P2 illustrates a frame of raw image captured by an image acquisition device after the structured light coded pattern shown is projected onto the actual surface of the anode carbon block. Because the surface of the anode carbon block is not an ideal plane but has a complex three-dimensional microstructure, the originally straight, parallel sinusoidal fringes are modulated on the surface, producing distortions that perfectly correspond to the surface height. The figure clearly shows the bending and twisting of the fringes, as well as potential local discontinuities caused by occlusion. This distortion information contains complete three-dimensional topographic data of the object's surface, serving as direct input for subsequent phase calculations and 3D reconstruction.

[0032] Figure 3Figure P3 shows the absolute phase map obtained after processing the original distorted image sequence using phase resolution and phase unrolling algorithms. This is a calculated data map whose macroscopic outline is consistent with the three-dimensional shape of the anode carbon block under test. The core information in the image is carried by the grayscale values ​​of the pixels: the grayscale value (from black to white) of each pixel uniquely and monotonically corresponds to an absolute phase value of that point in three-dimensional space. As shown, the grayscale of the entire surface exhibits a continuous and smooth gradient distribution without jumps or periodic repetitions, indicating that phase blurring has been successfully eliminated. This absolute phase map provides unambiguous and accurate input data for subsequent three-dimensional coordinate reconstruction.

[0033] Figure 3 Figure P4 illustrates the final 3D point cloud data output by the optical data acquisition module of this invention. This figure, from a 3D perspective with viewpoint, shows the surface model of the anode carbon block obtained after processing the absolute phase map using a 3D reconstruction algorithm. This model is not composed of traditional lines or surfaces, but rather a large number of discrete points with precise coordinates (x, y, z) in 3D space. This set of points forms the characteristic 3D point cloud, accurately reproducing the true 3D geometry of the anode carbon block surface, including its macroscopic undulations and microscopic rough texture. This high-precision 3D point cloud data will serve as the core data foundation for subsequent extraction of optical feature vectors and the final estimation of porosity distribution.

[0034] This embodiment aims to elaborate on the three core functional modules of the real-time porosity estimation system for low-porosity anode carbon blocks. First, the optical data acquisition module, characterized by using structured light 3D scanning technology to non-contactly acquire high-precision 3D point cloud data of the surface of the anode carbon block under test, serving as the raw input for characterizing its microstructure. Second, the physical property acquisition module, characterized by using tunable diode laser absorption spectroscopy (TDLAS) technology to accurately measure gas transport characteristic parameters that directly reflect the connectivity of the internal pore network at specified points, serving as the "ground truth" for model calibration. Third, the calibration decision module, characterized by intelligently determining the points with the highest information value based on the global optical microstructure data, guiding the physical property acquisition module in its measurements.

[0035] It should be further explained that the mapping model referred to in this invention refers to a core computing unit configured to receive an "optical microscopic topography data map" as input and output a "prediction result pair". The "prediction result pair" is a data structure comprising two parts:

[0036] The first part is the "predicted mean map," where each pixel value represents a predicted value of porosity at the corresponding location. In the context of this invention, this "predicted mean map" is referred to as the "initial porosity distribution map."

[0037] The second part is the "model uncertainty map," where each pixel value quantifies the degree of uncertainty the model has about its corresponding predicted value. In the context of this invention, this "model uncertainty map" is normalized and then weighted and fused with the feature novelty map, resulting in what is referred to as the "information value map."

[0038] In a preferred embodiment, if the mapping model is a Gaussian process regression (GPR) model, then its output posterior predicted mean map is the "predicted mean map," and its output posterior predicted variance map is the "model uncertainty map." In another embodiment, if the model is a deep neural network (DNN) that applies uncertainty quantification techniques such as Monte Carlo dropout, then the mean map of multiple forward propagation results is the "predicted mean map," and the standard deviation map of the results is the "model uncertainty map." Before starting the online optimization process of this invention, an initial mapping model with basic predictive capabilities needs to be constructed through offline training. This process is detailed below:

[0039] An initial training dataset for supervised learning is constructed. This process involves selecting at least 10 anode carbon blocks from different production batches from historical production data to ensure data diversity. For each sample block, a full-surface scan is performed using a high-precision 3D optical profilometer, and at least 100 sampling points with significant differences in optical features are extracted. Subsequently, for each of these sampling points, precise physical perforation measurements are performed using a gas permeameter to obtain its true gas transport coefficient value. This constitutes an initial dataset containing at least 1000 "optical permeability index-gas transport coefficient" data pairs. This dataset is then randomly divided into training, validation, and test sets at a ratio of 80%, 10%, and 10%, respectively.

[0040] In a preferred embodiment, if a Gaussian process (GPR) model is used, the training steps are as follows:

[0041] The quadratic exponential kernel function is chosen as the covariance function of the GPR. With maximizing the marginal log-likelihood as the optimization objective, the hyperparameters, such as the length scale and signal variance, of the kernel function are automatically solved and determined using numerical optimization algorithms such as gradient descent on the training set data. The trained GPR model is thus defined by the optimized hyperparameters and the complete training dataset.

[0042] In another embodiment, if a deep neural network (DNN) model is used, the training steps are as follows:

[0043] Define a fully connected neural network architecture consisting of an input layer matching the dimension of the optical permeability index, three hidden layers each containing 64 neurons (using ReLU as the activation function), and a linear output layer that outputs a single gas transport coefficient value. The mean squared error is used as the loss function, the Adam optimizer is selected, and the initial learning rate is set to 0.001. The network is batch trained using the training set, and the model's loss is evaluated on the validation set after each training epoch. If the validation set loss does not decrease for 10 consecutive epochs, an early stopping mechanism is triggered to terminate training. The network weights and bias parameters that perform best on the validation set during training are saved. After training, the performance of the initial model is finally evaluated using a reserved test set. The initial model is considered acceptable only if its coefficient of determination R² on the test set is higher than 0.9. The initial model is also considered acceptable only if its predicted mean plot on the test set has a coefficient of determination R² higher than 0.9, and its output model uncertainty plot shows a significant positive correlation with the actual prediction error distribution. The qualified model parameters will be serialized into a file that defines a computational entity capable of generating the aforementioned "prediction pairs" and deployed to the online system as the starting point for all subsequent incremental optimizations.

[0044] Further explanation: The optical data acquisition module is specifically used for:

[0045] Structured light 3D scanning technology is used to project a pre-coded structured light pattern onto the surface of the anode carbon block under test; and,

[0046] Images of structured light patterns modulated on the surface of the anode carbon block under test are acquired and demodulated to reconstruct three-dimensional point cloud data of the surface of the anode carbon block under test as optical micromorphological data.

[0047] In a preferred embodiment, the structured light 3D scanning technology uses a digital light processing (DLP) projector to project a high-frequency sinusoidal fringe pattern and simultaneously acquires images using an industrial camera. The phase principal value is demodulated using a multi-step phase-shifting algorithm, and an absolute phase map is obtained using a time-phase unfolding algorithm. Finally, combined with pre-calibrated system intrinsic and extrinsic parameters, a surface 3D point cloud with micrometer-level precision is reconstructed. All configurable operating parameters, including fringe frequency, projection brightness, and camera exposure time, are predefined and stored in an Excel spreadsheet file.

[0048] Further explanation: The physical property acquisition module is specifically used for:

[0049] At the calibration point, tunable diode laser absorption spectroscopy was used to measure the transient curve of the concentration of the probe gas after it penetrated the carbon anode block under test as a function of time; and,

[0050] The parameters characterizing the connectivity of the internal pore network of the anode carbon block under test are obtained based on transient curve calculations and used as gas transport characteristic data.

[0051] Specifically, this is achieved through an integrated measuring head. The measuring head seals the calibration point, applies a step pressure of CH4 probe gas to one side, and emits a laser beam of a specific wavelength through the escaping gas region from the other side. By analyzing the attenuation of the laser intensity, the gas concentration is calculated in real time. The entire concentration breakthrough curve is recorded. Gas transport characteristic data, specifically the gas breakthrough time or the gas transport coefficient obtained by fitting a physical model, are also provided. All relevant parameters, including probe gas type, applied pressure, laser wavelength, and data acquisition frequency, are configured through the aforementioned spreadsheet file. This invention discloses a high-precision, non-destructive physical measurement method for obtaining "ground truth" data. It miniaturizes and accelerates macroscopic and time-consuming permeability experiments, providing data that directly reflects the connectivity of internal pores.

[0052] Further explanation: The calibration decision module is specifically used for:

[0053] Based on the initial porosity distribution map, an information value map characterizing the uncertainty of model inference is calculated; and...

[0054] Select the peak point or centroid position in the information value map as the coordinates of the optimal calibration point command.

[0055] The calculation of information value, and its specific algorithmic logic, can be configured as one of several strategies and stored in the aforementioned spreadsheet file. For the "uncertainty sampling" strategy: if the mapping model is a Gaussian process regression model, then the information value map is the posterior variance map of the model output; if it is a neural network model, then it is approximated by the variance of the prediction results from multiple forward propagations using Dropout. Each time, the point with the largest "entropy reduction" in the global model's cognition is selected for calibration. The performance of this implementation is quantified by the calibration information gain factor, defined as the ratio of the improvement in model accuracy between the intelligent selection strategy and the random selection strategy under the same number of calibrations.

[0056] This limitation discloses that the decision-making mechanism of the present invention transforms from a traditional random or fixed sampling mode to a data-driven, dynamic, and precise deployment aimed at maximizing information gain. This improves the utilization efficiency of scarce resources and achieves the fastest improvement in global model performance with the fewest calibrations.

[0057] Further explanation: The steps for calculating the information value map in the calibration decision module specifically include:

[0058] Step a): Calculate a model uncertainty diagram based on the initial porosity distribution map;

[0059] Step b): Extract the optical feature vector of each pixel from the optical micromorphology data and calculate a feature novelty map. The numerical value of the feature novelty map represents the degree of deviation of the optical feature vector of each pixel from a set of historical feature vectors.

[0060] Step c): After normalizing the model uncertainty map and the feature novelty map, perform weighted fusion to generate a hybrid information value map as the information value map.

[0061] Further explanation: When utilizing optical microstructure data, the calibration decision module is specifically used for:

[0062] Based on the three-dimensional point cloud information in optical micro-topography data, the geometric features and spatial distribution features of pores in local areas are identified and quantified.

[0063] Based on the pore geometry and spatial distribution characteristics, a one-dimensional optical permeability index is calculated using a pre-calibrated weighted model; and,

[0064] The optical permeability index was used as input to the Gaussian process regression model to generate an initial porosity distribution map and a model uncertainty map.

[0065] The following is a detailed description of the implementation of the above content: This embodiment is configured to acquire optical microstructure data of the surface of the anode carbon block to be tested. In a specific embodiment, a structured light three-dimensional scanning subsystem is included;

[0066] The structured light encoded pattern, denoted as Ppattern, represents a two-dimensional image pattern projected onto the surface of the object under test for encoding depth information. Specifically, an N-step phase-shifting sinusoidal fringe pattern is used, where N is an integer greater than or equal to 3. In a preferred embodiment, N is set to 4, i.e., a four-step phase-shifting method is used. The spatial frequency and fringe direction parameters of the pattern are read from a locally stored Excel spreadsheet configuration file.

[0067] The original distorted image sequence is denoted as I. raw (n); represents the nth phase-shifted structured light pattern image modulated by the surface of the anode carbon block, captured by an industrial camera. Specifically, it is acquired synchronously by an industrial camera with a resolution of 1920×1080 pixels, as each step of the pattern is projected by a DLP projector.

[0068] An absolute phase map, denoted as Φabs, represents a two-dimensional image with the same resolution as the camera image. The grayscale value of each pixel directly corresponds to the unique phase value of that point in the scene within the projector coordinate system. This is determined through the following calculation model:

[0069] The input is the original distorted image sequence I raw(n), where n ranges from 1 to N. Essentially, it is a modulated sinusoidal light intensity signal carrying depth information. Before performing phase calculation, the original distorted image sequence I... raw Each frame in (n) is applied with a Gaussian filter to suppress random Gaussian noise introduced by the image sensor. The Gaussian filter is a two-dimensional Gaussian filter with a kernel size of 3×3, and its core parameter, the standard deviation, determines the smoothing strength of the filter. In a preferred embodiment of the invention, the standard deviation is set to 0.8. This value is determined experimentally by applying Gaussian filters with different standard deviations to a series of standard sample images and evaluating the signal-to-noise ratio and spatial resolution of the subsequent phase demodulation results. The value of 0.8 is an empirical value that achieves the best balance between noise suppression and phase detail preservation. In other embodiments, the standard deviation is preferably between 0.5 and 1.5; values ​​outside this range may result in insufficient denoising or excessive loss of detail.

[0070] For each pixel (u,v) in the image, the processing module performs the following calculations: It obtains the grayscale value sequence of the pixel (u,v) across N steps of the image; using the arctangent function, it calculates a wrapping phase value Φ with a value range between negative π and positive π based on this grayscale value sequence. wrap (u,v). The idea for this computation comes from the Discrete Fourier Transform in the field of signal processing.

[0071] The processing module employs time-phase unrolling technology, projecting multiple sets of coded patterns at different frequencies and synthesizing the demodulation results to determine a unique period ordinal k(u,v) for each pixel. The resulting wrapper phase value Φ is then used to... wrap The absolute phase value Φabs(u,v) is obtained by linearly combining the period ordinal k(u,v) with the period ordinal k(u,v), specifically by multiplying the period ordinal k(u,v) by 2π and adding it to the wrapped phase value. This invention preferably employs a time-phase unfolding method. This method involves additionally projecting and acquiring multiple sets of phase-shift fringe patterns at different spatial frequencies (frequency ratio 1:8:64). The processing module calculates the wrapped phase map at each frequency, and then, starting from the lowest frequency, unwrapped phase map, progressively and pixel-by-pixel corrects the periodic blurring of the higher frequency wrapped phase maps, ultimately obtaining a blur-free absolute phase map Φabs across the entire field.

[0072] 3D point cloud data, denoted as C pointcloud It represents an ordered dataset containing M×N three-dimensional coordinate vectors (x, y, z), describing the three-dimensional geometric morphology of the anode carbon block surface, and serving as the final output "optical microstructure data". Specifically, it is determined through the following computational model:

[0073] The input is the absolute phase map Φabs, and the pre-calibrated system intrinsic and extrinsic parameters (including the camera intrinsic parameter matrix K).cam Projector intrinsic parameter matrix K proj , and the rotation matrix R and translation vector T between them).

[0074] The objective of the experiment is to determine the internal distortion parameters, focal length, principal point coordinates, and relative rotation and translation between the camera and projector in three-dimensional space.

[0075] The experimental subject was a high-precision planar checkerboard calibration board with 10×7 corner points and each square having a side length of 5 mm. The optical data acquisition module included a camera and a projector. The experiment was conducted in a dark room with stable lighting and no stray light interference. Specifically, the calibration board was placed within the measurement field of view to ensure a complete and clear image. Two sets of imaging data were acquired from at least 15 different random poses (including different distances, tilt angles, and rotation angles): the first set was images of the calibration board directly captured by the camera; the second set was images captured by the camera after the same checkerboard pattern was projected onto the calibration board by the projector. For all acquired images, a corner detection algorithm (Harris or Shi-Tomasi algorithm) was used to extract the two-dimensional image coordinates of all corner points with sub-pixel accuracy.

[0076] Using the corner coordinates in the first set of images, and based on Zhang Zhengyou's camera calibration method, the camera's intrinsic parameter matrix K is solved through an iterative optimization process. cam And distortion coefficients. Treat the projector as an inverse camera. Using the second set of images, combined with the known camera parameters, solve for the projector's intrinsic parameter matrix K. proj By combining the two sets of data, the rotation matrix R and translation vector T between the camera coordinate system and the projector coordinate system are calculated. The final parameters are stored in the configuration file. Verification is performed using backprojection error; that is, using the calibrated parameters, corner points in the 3D world coordinate system are backprojected back onto the image plane, and the pixel distance between them and the actually detected image corner points is calculated, ensuring that the average backprojection error is less than 0.1 pixels. Before running, a high-precision checkerboard calibration board is photographed. This calculation logic is based on a pinhole camera model and the triangulation principle in stereo vision. The process of determining the 3D coordinates (x, y, z) of each effective pixel (u, v) in the absolute phase map Φabs is as follows:

[0077] The input quantities include the image coordinates (u,v) of the pixel, its corresponding absolute phase value Φabs(u,v), and the system intrinsic and extrinsic parameters (K) obtained from the aforementioned experimental calibration. cam ,K projBased on the camera model and pixel coordinates (u,v), a 3D spatial ray originating from the camera's optical center and passing through the pixel is determined. Based on the projector model and the absolute phase value Φabs(u,v), a 3D spatial plane originating from the projector's optical center and defined by this phase value is determined. In the same world coordinate system, the unique intersection point of the above 3D spatial ray and the 3D spatial plane is calculated, and the 3D coordinates of this intersection point are (x,y,z); finally, the 3D point cloud data is output.

[0078] Further explanation: The complete workflow of the optical data acquisition module is as follows: Load runtime parameters from the configuration file. Drive the DLP projector to sequentially project an N-step phase-shift structured light coded pattern. During the projection of each pattern step, simultaneously trigger the industrial camera to acquire a frame of raw distorted image. After acquisition, execute the "Absolute Phase Map Determination Model" to calculate the absolute phase map. Finally, execute the "3D Point Cloud Data Determination Model" to reconstruct the final 3D point cloud data using the absolute phase map and pre-calibration parameters, and output it as "Optical Microstructure Data" to subsequent modules.

[0079] Furthermore, this embodiment is configured to acquire gas transport characteristic data of the anode carbon block at a calibration point specified by the "calibration decision module".

[0080] The optical permeability index is denoted as OPI. The calculation logic is as follows: the input is a local neighborhood in the 3D point cloud data; in this embodiment, it is an 11×11 pixel window centered on the current point. The average Z-coordinate of all points within this neighborhood is calculated and used as the local reference plane height H. base Iterate through each pixel in the neighborhood; if its Z-coordinate is lower than the height H of the reference plane... base A preset depth threshold (e.g., below H_base by 3 standard deviations) will mark the point as a "pore pixel"; in this embodiment, the depth threshold is below the reference plane height H. base The three standard deviations of the data are used to divide all marked "pore pixels" into N1 independent pore regions using a connected component analysis algorithm. For each independent pore region i, its projected area Ai and average depth Di are calculated. Based on the above quantification results, the following three statistical indicators characterizing the overall permeability potential of the region are calculated:

[0081] Pore ​​area density is denoted as ρ area : Calculate the ratio of the total area of ​​all pore regions to the total area of ​​the neighborhood.

[0082] The weighted average depth is denoted as D. avg Calculate the weighted average of all pore depths, using the area of ​​each pore as the weight.

[0083] Pore ​​connectivity factor denoted as C factorCalculate the number of independent pore regions N1 within a neighborhood. Given a similar total pore area, a smaller number of independent pores indicates a greater likelihood of pores merging into large connected regions, resulting in stronger permeability. The pore connectivity factor is defined as the ratio of the total pore area to the number of independent pores N1.

[0084] The Optical Permeability Index (OPI) is obtained by linearly weighting and summing the three statistical indicators mentioned above; its calculation logic is as follows: The pore area density ρ... area Multiply by the first weighting coefficient W1; then calculate the weighted average depth D. avg Multiply by the second weighting coefficient W2; then multiply by the pore connectivity factor C. factor Multiply by the third weighting coefficient W3; add these three products together to get the final OPI value.

[0085] The experimental calibration method for the OPI model weight coefficients (W1, W2, W3) is as follows: Prepare at least 20 standard anode carbon block samples with different porosities. For each sample, select multiple points uniformly on its surface, and perform operations of the optical data acquisition module and the physical property acquisition module respectively to obtain a set of regional pore network characteristics (ρ). area D avg C factor ) and the corresponding actual gas transport coefficient K gas The data pairs.

[0086] With K gas As the dependent variable, with (ρ) area D avg C factor Using K as the independent variable, we apply multiple linear regression analysis to find the value that makes the predicted value consistent with the true value. gas The optimal set of weight coefficients that minimizes the sum of squared residuals among the values ​​is the final set of (W1, W2, W3).

[0087] The concentration transient curve, denoted as Ctransient, represents one-dimensional time series data, recording the change in probe gas concentration in the measurement chamber over time after the application of a step pressure. It is specifically obtained through real-time measurement using the TDLAS subsystem.

[0088] It should be noted that, in a preferred embodiment of the present invention, the core of the physical characteristic acquisition module is a tunable diode laser absorption spectroscopy (TDLAS) subsystem. This section will explain in detail the physical integration structure, control logic, and data interface of this subsystem, as well as the specific signal processing flow for this application.

[0089] To achieve precise, in-situ measurement of any specified point on the surface of the anode carbon block, the optical and gas path components of the TDLAS subsystem are integrated into a compact, specially designed measuring head driven by a multi-axis industrial robotic arm.

[0090] Measuring head body and mounting interface: The measuring head body is machined from a single piece of 316L stainless steel to ensure structural rigidity and chemical inertness. Its top is designed with a standard flange interface for rigid connection to the end effector of the industrial robotic arm.

[0091] Double-layer flexible gas-tight mechanism: Considering the inherent roughness and porosity of the anode carbon block surface, in order to ensure the accurate definition and airtightness of the measurement area, this invention designs a double-layer sealing structure.

[0092] Outer flexible pre-sealing ring: Made of highly flexible silicone rubber material, its function is to adapt to the macroscopic unevenness of the surface when the measuring head contacts the carbon block surface, forming a large pre-sealing area to prevent the intrusion of ambient gases.

[0093] Inner rigid precision sealing ring: Made of corrosion-resistant and high-hardness fluororubber, its inner diameter precisely defines the effective area for measuring gas transport characteristics, with a diameter of 5 mm. This inner sealing ring, under a constant pressure of 50 Newtons applied by a miniature pneumatic actuator and set via a precision pressure reducing valve, can embed into the micropores on the carbon block surface, forming a reliable high-pressure airtight seal.

[0094] The measuring head body has a miniature gas path machined inside through precision drilling. One path is a central injection channel, used to inject probe gas (including but not limited to methane) into the center of the measuring area; the other path is an annular discharge channel, evenly distributed on the outside of the inner sealing ring, used to quickly discharge gas from the chamber. Both channels are connected to an external electrically controlled high-speed solenoid valve assembly via high-pressure hoses.

[0095] To increase the effective optical path between the laser and the gas within a limited measurement head volume, thereby improving measurement sensitivity, a miniature folded optical path chamber is designed inside the measurement head. The laser beam enters through a sealed sapphire window, is reflected twice by gold-plated mirrors positioned above the measurement area (approximately 1 mm above the carbon block surface), forming a "V"-shaped optical path, and finally exits through another sapphire window to the photodetector. This design ensures that the laser beam passes through the probe gas escaping from the carbon block surface twice, effectively doubling the optical path. The TDLAS subsystem includes a separate microcontroller unit (MCU) responsible for performing precise timing control and interacting with the system's central processing unit (CPU). The MCU establishes a TCP / IP connection with the CPU via an industrial Ethernet interface. It is configured to listen on a specific port and parse a JSON-formatted instruction string. This string contains the command to execute the measurement, a unique request ID, and the physical world coordinates calculated by the calibration decision module.

[0096] After the MCU receives and verifies the above instructions, it will strictly follow the following preset automation timing sequence to drive the relevant hardware through the I / O ports and execute the complete measurement process:

[0097] The MCU sends a confirmation response to the host computer. The MCU forwards the target coordinates to the robotic arm controller via the bus, and the robotic arm drives the measuring head to move to the designated position. Once the robotic arm is in position, the MCU activates the micro-pneumatic actuator to apply pressure and complete the seal. Immediately, the MCU triggers the TDLAS laser and the data acquisition card (DAQ) to acquire a 200-millisecond signal as the background baseline signal without introducing probe gas. The MCU simultaneously executes two actions with microsecond precision: ① opening the probe gas injection valve; ② instructing the DAQ to begin real-time recording of the photodetector's output voltage signal at a high sampling rate (10kHz). This process records the raw electrical signal of the concentration transient process. After continuous injection and recording for 500 milliseconds, the MCU closes the injection valve and stops data acquisition. The MCU opens the discharge valve to purge the measurement chamber and resets the measuring head. Simultaneously, it encapsulates the recorded time-series data along with metadata such as the request ID and coordinates into a JSON-formatted response packet and sends it back to the central processing unit via a TCP / IP connection.

[0098] The raw voltage signal sequence recorded by DAQ is transformed into a physically meaningful concentration transient curve for subsequent analysis. This is achieved through the following targeted signal processing: the average value of the acquired background baseline signal is subtracted point-by-point from the entire transient voltage signal sequence to eliminate the effects of system DC bias and slow drift. To suppress noise and improve the signal-to-noise ratio, wavelength modulation spectroscopy is preferably employed. Specifically, a high-frequency 5kHz sine wave is used to slightly modulate the DC current driving the diode laser, causing the laser wavelength to sweep near the target absorption spectral line. The photodetector signal is then fed into a digital lock-in amplifier module, configured to demodulate the harmonic components (2f signal) at twice the modulation frequency of 10kHz. Theoretically, the peak amplitude of this 2f signal is proportional to the target gas concentration over a wide range. Before deployment, a series of standard gases of known concentrations (precisely proportioned by a gas mixer) are introduced into the measuring head, and the corresponding stable 2f signal peak amplitudes are measured and recorded. Based on these data points, a calibration lookup table or polynomial fitting function is established to accurately and non-linearly convert any measured 2f signal amplitude value into a gas concentration value in SI units. The time series data processed through all the above steps constitutes the final concentration transient curve Ctransient, which can be directly used for fitting the physical model. The horizontal axis of this curve represents time (in seconds), and the vertical axis represents the calibrated probe gas concentration (in ppm).

[0099] The gas transport coefficient, denoted as K. gas This is a scalar value that quantitatively characterizes the degree to which the porous medium inside the anode carbon block hinders the flow of selected gas; a smaller value indicates greater density. It serves as the final output "gas transport characteristic data." Specifically, it is determined through nonlinear fitting of the concentration transient curve, as detailed below.

[0100] The computational logic is based on the one-dimensional convection-dispersion equation in porous media seepage theory, which describes the temporal and spatial variation of fluid concentration during transport in porous media. The input is the measured transient concentration curve Ctransient. The processing module contains a built-in numerical solution to the convection-dispersion equation. This numerical solution is a function whose input is time t and a set of model parameters (including the gas transport coefficient K to be determined). gas The effective porosity (Kxd) is used, and the output is the theoretical gas concentration. A Levenberg-Marquardt nonlinear optimization algorithm is employed, with the gas transport coefficient K... gas Let Kxd be the variables to be optimized, and the root mean square error between the measured concentration transient curve Ctransient and the theoretical curve be used as the objective function. Starting from an initial guess, the objective function is iteratively calculated to affect the gas transport coefficient K. gas The gradient of Kxd is used to continuously adjust K. gas The values ​​of Kxd and K are used until the objective function converges to its minimum. The value of K at convergence is then returned. gas The value is used as the final result. In this embodiment, the initial guess value and convergence threshold of the optimization algorithm are stored as configuration parameters in an external file.

[0101] It should be noted that upon receiving a command containing the target coordinates, the robotic arm drives the integrated measuring head to move to the designated point on the surface of the anode carbon block and seals it. The gas control system applies a step pressure probe gas. The TDLAS subsystem begins real-time recording of the concentration transient curve Ctransient. After data acquisition, the processing module executes the "deterministic model for gas transport coefficients" to calculate K. gas The value is then output as "gas transport characteristics data".

[0102] The model uncertainty map, denoted as Muc, represents a two-dimensional map of the same size as the initial porosity distribution map. Its pixel values ​​characterize the degree of uncertainty in the model's prediction of porosity at that point. A nonlinear mapping relationship is constructed from high-dimensional optical feature vectors to scalar gas transport coefficients. This invention preferably employs a Gaussian process regression (GPR) model, which provides the variance of the prediction, directly and quantitatively characterizing the model's uncertainty regarding that prediction. The training dataset consists of a series of data pairs (V... feat ,K gasThe data are accumulated by performing multiple "optical data acquisition" and "physical property acquisition" operations on the sample. The trained GPR model is deployed in the system. When an entire optical feature vector map of the sample is input, the model outputs two maps in parallel: one is the predicted mean map, i.e., the initial porosity distribution map; the other is the predicted variance map, which is directly used as the model uncertainty map Munc.

[0103] Optical eigenvector, denoted as V feat This represents a numerical vector extracted from local optical microscopic morphology data, capable of characterizing multi-dimensional features such as texture and roughness of the region. Its original data is a local rectangular neighborhood centered on the point to be calculated in the 3D point cloud data. In this embodiment, the size of the neighborhood defines the spatial scale of feature extraction and should match the typical microstructural dimensions (average pore size or particle diameter) of the surface of the anode carbon block to be tested. In this embodiment, for anode carbon blocks with an average pore size in the range of 50 to 200 micrometers, if the lateral resolution of the 3D point cloud data is 10 micrometers per pixel, the size of the neighborhood is preferably 11×11 pixels. In other embodiments, this size can be adjusted within the range of 7×7 to 21×21 pixels to adapt to the surface morphology analysis needs at different scales. This size is determined in advance by performing metallographic microscopy on the sample to determine the statistical size distribution of its key surface features, and then selecting a window size that can completely cover one or several typical features. For each local neighborhood, the processing module calculates the following set of statistics and concatenates them into a multi-dimensional optical feature vector V. feat :

[0104] Calculate the standard deviation and skewness of the Z-coordinates of all points within the neighborhood: Fit a quadratic surface to the neighborhood, then calculate the mean and standard deviation of the angles between the normal vectors of all points within the neighborhood and the normal vector of the center point. Calculate the ratio of the principal curvatures within the neighborhood, and the mean of the Gaussian curvature. These statistics constitute the optical eigenvector V. feat Each dimension of the vector has a fixed number of dimensions, ensuring that effective mathematical operations can be performed between vectors.

[0105] Historical characteristic centroid, denoted as C hist This represents a vector with the same dimension as the optical eigenvectors, representing the average center of the optical eigenvectors of all historically calibrated points. The system maintains a V-shaped array of historically calibrated points. feat Set, C hist It is the arithmetic mean of all vectors in the set.

[0106] The feature novelty map, denoted as Mnovel, is a two-dimensional image where pixel values ​​characterize the rarity of the optical feature at a given point. It is determined using the following computational model: Input the optical feature vector V for each pixel. featAnd historical characteristics centroid C hist For each pixel, calculate its V. feat With C hist The Euclidean distance between the points is calculated. This calculated distance value is used as the value of that pixel in Mnovel. A larger distance indicates a more novel optical feature at that point. The feature novelty map Mnovel is then output.

[0107] Hybrid information value map; denoted as Mhybrid; represents the comprehensive value map that integrates uncertainty and novelty information ultimately used for decision-making. It is determined through the following computational model:

[0108] The input model consists of an uncertainty map (Munc), a feature novelty map (Mnovel), and two weight coefficients read from the configuration file: uncertainty weight (Wunc) and novelty weight (Wnov). The processing module performs min-max normalization on both Muc and Mnovel, linearly mapping all their values ​​to a uniform interval of [0,1], resulting in the normalized Muc. norm and Mnovel norm .

[0109] For each pixel, perform the following calculation: place the point in Munc norm The value in the value is multiplied by the uncertainty weight Wunc, and the point is then plotted in Mnovel. norm The value in the value is multiplied by the novelty weight Wnov, and then the two products are added together to obtain the hybrid information value of that point; finally, the hybrid information value map Mhybrid is output.

[0110] The uncertainty weight Wunc and the novelty weight Wnov are determined through configuration. Specifically, Wunc represents the preference for "exploitation," i.e., prioritizing in-depth exploration of regions where the model's understanding is ambiguous; Wnov represents the preference for "exploration," i.e., prioritizing the exploration of entirely new feature regions that the model has not yet seen. In this invention, these two weights are constrained, and their sum is a constant, preferably 1.

[0111] During the system's cold start or initial learning phase, the model's understanding is severely limited, making the exploration of the feature space the primary task. In this phase, the optimal parameter configuration is: Wunc=0.3, Wnov=0.7.

[0112] During the system's stable learning or model fine-tuning phase, the model has already acquired a certain level of generalization ability, and the main task is to eliminate local uncertainties. In this phase, the optimal parameter configuration is: Wunc = 0.8, Wnov = 0.2.

[0113] The core of this embodiment lies in the "hybrid information value" fusion mechanism. Conventional techniques rely solely on Munc. When the model's overall understanding of a certain type of feature is insufficient, it can lead to large areas of high values ​​in the Munc. In this case, if only peak points are selected, repeatedly selecting points with similar features results in "repetitive learning," which is inefficient.

[0114] This invention addresses the aforementioned limitations by introducing Mnovel as a second decision dimension. Mnovel represents the exploration of the feature space, while Munc represents the utilization of known, ambiguous regions. By weighted fusion of the two, a dynamic balance can be achieved between "utilization" and "exploration." When Wunc > Wnov, the model tends to delve deeper into known, uncertain regions; when Wnov > Wunc, the model tends to explore new feature regions never seen before. This balancing strategy originates from the "exploration-utilization dilemma" in reinforcement learning, ensuring that the model can learn and converge more quickly and comprehensively.

[0115] The calibration decision module receives the initial porosity distribution map and the corresponding optical microstructure data. It executes two computational branches in parallel: Branch 1: Calculates the uncertainty map Munc of the model. Branch 2: Extracts the optical feature vector V for each point. feat And execute the "deterministic model of feature novelty map" to calculate Mnovel.

[0116] The "deterministic model for the hybrid information value map" is executed, fusing Munc and Mnovel to generate the final hybrid information value map Mhybrid. A global maximum point is searched on Mhybrid, and its physical coordinates are encapsulated into an "optimal calibration point instruction." This instruction is output to the execution mechanism of the physical characteristic acquisition module. Simultaneously, the V of this calibration point is... feat Update to the historical feature vector set and recalculate C. hist This is to prepare for the next decision.

[0117] By normalizing all graphs (Munc, Mnovel) involved in the fusion calculation to the [0,1] interval, dimensional consistency is ensured. Under boundary conditions, when the system model is very mature and Munc is generally low, decision-making will be mainly driven by Mnovel, prompting the system to verify some rare operating conditions; conversely, when the morphology of the sample to be tested is very uniform and Mnovel is generally low, decision-making will be mainly driven by Munc, focusing on solving the remaining cognitive blind spots of the model. This mechanism ensures that the system can operate stably and reasonably at different stages.

[0118] To decouple the core algorithm from specific application strategies and ensure the configurability and ease of debugging of the technical solution, all configurable operating parameters are predefined and stored in a structured external data carrier in the specific implementation path of this invention. Preferably, this data carrier is a locally stored Excel spreadsheet file. This file contains multiple worksheets. During the initialization phase, the system reads this file and loads all parameters into the working memory. Subsequent algorithm modules will directly query and retrieve the required configuration parameters from memory, thereby achieving the stability of the algorithm logic and the flexibility of the application strategy.

[0119] Further explanation: The model optimization module is specifically used for:

[0120] The optical microstructure data determined by the optimal calibration point command and the corresponding gas transport characteristic data are used to form a new training data pair; and,

[0121] By employing transfer learning or Bayesian update algorithms, the mapping model can be fine-tuned online using new training data pairs.

[0122] Further explanation: The results generation module is specifically used for:

[0123] Based on the inherent uncertainty parameters of the incrementally optimized mapping model, or based on the spatial distance between each point in the initial porosity distribution map and the historical calibration points, the confidence score corresponding to each point in the calibrated porosity distribution map is calculated to form an estimated confidence distribution map.

[0124] Further explanation: The steps for calculating the confidence score in the results generation module specifically include:

[0125] Based on the inherent uncertainty parameters of the mapping model, calculate an intrinsic confidence map of the model;

[0126] Based on the time-decayed spatial distance between each point and the historical calibration points, a calibration proximity confidence map is calculated.

[0127] Calculate a scalar value for the confidence of a model state based on the residuals of the training data pairs used in the most recent incremental optimization of the model.

[0128] The model's intrinsic confidence map and the calibrated neighbor confidence map are weighted and fused, and the fusion result is modulated using the model state confidence to generate an estimated confidence distribution map.

[0129] The following are specific implementation instructions for the above content:

[0130] The new training data pair, denoted as Dnew, represents the optical permeability index (OPI) and its corresponding physically measured gas transport coefficient (K) at the optimal calibration point. gasThis constitutes a binary tuple, which is the information source for model optimization. It is triggered by the output command of the calibration decision module and combined by the optical and physical property acquisition module after collaborative measurement.

[0131] The freeze layer mask, denoted as Mfreeze, is used when mapping a deep neural network. It is a Boolean vector that specifies which layers' weights remain unchanged during fine-tuning. This freeze layer mask is set to freeze all layers except the last N2 fully connected layers, where N2 is a configurable integer read from a locally stored Excel spreadsheet configuration file.

[0132] Fine-tuning the learning rate, denoted as η finetune This value is used when the mapping model is a deep neural network and controls the step size of model weight updates during fine-tuning. It is set to a positive real number much smaller than the initial training learning rate; in this embodiment, it is 1% of the initial learning rate. The specific value is read from the configuration file mentioned above.

[0133] Online fine-tuning computational models for deep neural networks, i.e., representations based on transfer learning:

[0134] Input new training data to Dnew, the current neural network model, the freeze layer mask Mfreeze, and the fine-tuned learning rate η. finetune The processing module sets the trainable attribute of the layers in the network that need to be frozen to false based on the freeze layer mask Mfreeze. An Adam optimizer is initialized, and its learning rate parameter is set to η. finetune Using the new training data pair Dnew as the training sample and the mean squared error as the loss function, a short training process is performed. In this embodiment, only 10 epochs are trained, and the number of epochs is configurable. During this process, only the weights of the layers that are not frozen are fine-tuned according to the gradient of the loss function; the output is a finely tuned neural network model with some updated weights.

[0135] This embodiment focuses on the online fine-tuning computation model of the Gaussian process model, specifically the characterization based on Bayesian updates:

[0136] The new training data pair Dnew is input. The current state of the Gaussian process model is defined by the historical training dataset and kernel function parameters. The processing module appends Dnew to the model's historical training dataset. Based on the updated training dataset, the model's covariance matrix is ​​recalculated. Numerical computation techniques such as block matrix inversion or Cholesky update are used to avoid completely recalculating the entire matrix. Based on the reconstructed covariance matrix, the model automatically updates the analytical expression of its posterior distribution. A Gaussian process model containing information on the new data points and updated posterior distribution is output. When the model optimization module is triggered, it receives Dnew. According to the model type specified in the configuration file, it selects to execute one of the two "computation models" mentioned above. After execution, the updated mapping model parameters overwrite the original model parameters, completing one incremental optimization loop, and waiting for the next trigger. This module is configured to use the updated mapping model to generate a corrected porosity distribution map and an estimated confidence distribution map that integrates multi-source information.

[0137] The intrinsic confidence map, denoted as Mmic, is a two-dimensional graph where pixel values ​​reflect the model's confidence in its predictions at each point. A Gaussian process regression (GPR) model is preferred, aiming to establish a mapping from the "optical permeability index" to the "gas transport coefficient." Based on Bayesian theory, the GPR model, for any given input point, outputs not only a predicted mean but also a complete predicted probability distribution. The variance of this distribution represents the uncertainty of the model's prediction at that point. During system operation, the trained GPR model outputs a posterior prediction variance value for each pixel on the surface under test (corresponding to the optical permeability index input). All these variance values ​​together constitute a "posterior prediction variance map." The determination of the intrinsic confidence map Mmic involves performing a point-by-point function transformation on the aforementioned "posterior prediction variance map." This transformation includes the following steps:

[0138] The "posterior prediction variance plot" is linearly normalized, mapping all its values ​​to the [0,1] interval. The maximum and minimum variance values ​​in the plot are found. Then, for any point in the plot, the variance value is calculated by dividing the minimum variance value by the maximum variance value. The normalized variance plot is then subtracted point by point from the value of 1. Higher variance corresponds to lower confidence, and the two are inversely related. After this transformation, in the final Mmic plot, a value of 1 represents the highest confidence level, and a value of 0 represents the lowest confidence level.

[0139] The calibration proximity confidence map, denoted as Mcc, is a two-dimensional graph where pixel values ​​reflect the confidence level of each point due to its proximity to historical accurate calibration points. It is determined through the following calculation model that, in industrial production environments, the "optical-physical" characteristic mapping relationship of anode carbon blocks undergoes a slow "process drift" due to factors such as equipment wear and tear and batch variations in raw materials. Therefore, recently obtained calibration data reflects the current true state of the system better than outdated calibration data, and its reference value is higher.

[0140] The input quantities are defined as: 1) the two-dimensional coordinates of the pixels whose confidence level is to be calculated; 2) a historical calibration database, which stores the two-dimensional coordinates of each historical calibration point and the timestamp when it was collected.

[0141] The algorithm iterates through each calibration point in the historical calibration database, calculating the time difference between the current system time and the timestamp of that point. This time difference is then input into a preset exponential decay function to obtain the "time decay weight" for that calibration point. This function's form is that the weight value decreases from 1 to 0 as the time difference increases. For the pixel to be calculated, the algorithm iterates through all historical calibration points, calculating the Euclidean distance between the pixel and each historical calibration point, and then divides this distance by the corresponding historical calibration point's "time decay weight" to obtain a series of "weighted distances." The minimum value among all the obtained "weighted distances" is selected as the final "proximity metric" for that pixel. This "proximity metric" is denoted as D1. min The input is a well-defined exponential decay function to calculate the final confidence score. This function takes the form: the confidence score equals the negative distance decay coefficient of the natural constant e multiplied by D1. min The "distance decay coefficient" converts the proximity metric into a confidence score over the [0,1] interval, calibrating the proximity confidence map Mcpc. The "time decay coefficient" and "distance decay coefficient" in the two exponential functions mentioned above are key internal parameters that determine the behavior of this model.

[0142] The model state reliability, denoted as Smst, is a scalar value ranging from [0,1], globally representing the "health status" or "reliability" of the current mapping model. It is determined through the following computational model:

[0143] Input the most recent training data pair Dnew used for model optimization. Using the unoptimized mapping model, predict the optical permeability index in Dnew to obtain a predicted K. gasThe absolute difference between the predicted value and the actual Kgas value in Dnew is calculated; this is the residual. This residual value is then input into an exponentially decaying function. This function is designed to output 1 when the residual is zero, and the output value rapidly decays towards 0 as the residual increases. The decay rate of the function is controlled by a configurable sensitivity coefficient. Finally, the model state confidence score Smst is output.

[0144] Mixed confidence score plot, denoted as M1 conf This represents the final output, an estimated confidence distribution map that incorporates information from multiple sources. It is determined using the following computational model: inputs Mmic, Mcpc, Smst, and two weighting coefficients read from the configuration file: the intrinsic confidence weight Wmic and the neighboring confidence weight Wcpc.

[0145] Ensure all values ​​of Mmic and Mcpc are within the range [0,1]. For each pixel, perform the following calculation: multiply the pixel's value in Mmic by Wmic, multiply the pixel's value in Mcpc by Wcpc, and then add the two products to obtain a base fusion confidence score. Multiply each pixel value in the resulting base fusion confidence score map by the model state confidence score Smst. Output the final mixture confidence score map M1. conf .

[0146] It should be further noted that the optimal values ​​for the time decay coefficient and the distance decay coefficient are determined through analysis of historical production data. Specifically, a historical dataset with a sufficiently long time span is collected. This dataset contains optical scan images arranged in chronological order and corresponding, post-hoc measured true porosity distribution maps.

[0147] Iterate through every parameter combination in the 2D mesh. For each parameter combination, perform a complete historical data playback simulation:

[0148] Following the chronological order of historical data, the online operation process of this invention is simulated point by point: prediction is performed, calibration points are selected, physical measurements are simulated, the model is updated, and the current time decay coefficient and distance decay coefficient are used to calculate the mixed confidence score map M1 for that moment. conf Simultaneously, the "true prediction error map" between the model prediction map and the actual porosity map at that moment is calculated. After each complete simulation, the error of all generated M1 over the entire simulation period is calculated. conf The overall Pearson correlation coefficient between the sequence and all corresponding "true prediction error map" sequences. The combination of "time decay coefficient and distance decay coefficient" that maximizes the above overall Pearson correlation coefficient is selected as the optimal parameters finally fixed in the system configuration file.

[0149] The core of this embodiment lies in the "hybrid confidence fusion mechanism." By weighted fusion of Mmic and Mcpc, the model's theoretical confidence and actual calibration evidence are complemented. Smst is introduced as a global modulation factor. Smst reflects the model's fitness level. If the residual from the most recent update is large, it indicates a serious conflict between the new data and the model's original understanding; Smst decreases, thus lowering the overall confidence of all points. This mechanism effectively suppresses "false high confidence" caused by anomalous data points contaminating the model, enhancing the robustness and reliability of confidence assessment.

[0150] The following detailed implementation description applies to the above content: In this embodiment, the confidence score range for any point in the "Estimated Confidence Distribution Chart" is [0,1];

[0151] When the confidence score approaches 1, it indicates that the system has a higher assessment of the reliability of the porosity prediction for that point. This state is caused by one or more of the following: (1) the optical characteristics of the point are within the model's cognitive range, and the model's inherent uncertainty is lower; (2) the point is spatially closer to a recent physical calibration point, and there is stronger support from "ground reality" data; (3) the point has not experienced drastic changes in recent times, and the model as a whole is in a highly reliable state. A confidence score approaching 1 is a clear signal to the user that the prediction result for that point is more trusted.

[0152] When the confidence score approaches 0, it indicates that the reliability assessment of the porosity prediction value of the system at that point is lower. This state reflects the risk identification capability of the present invention and is caused by one or more of the following situations: (1) The optical features of the point belong to samples that the model has not seen, triggering model extrapolation and increasing the inherent uncertainty; (2) The further the point is from any historical calibration point and the longer the time since the nearest calibration point, the higher the risk of process drift; (3) Incremental optimization has been performed through new data points, and the larger the prediction residual, the more the confidence of the model state drops sharply, and the system performs a global active downgrade of the reliability of all its predictions. A confidence score approaching 0 is a strong warning to the user: the prediction result at that point is less reliable.

[0153] When other parameters remain constant, an increase in the Mmic value at any point will lead to a monotonically increasing final confidence score. This is a positive correlation. Mmic is a direct quantification of the predictive power of a mapping model based on its statistical principles. Models naturally have higher confidence in areas with dense data and familiar features; this positive correlation design accurately uses the model as a fundamental component of confidence assessment.

[0154] When other parameters remain constant, an increase in the Mcpc value at any point leads to a monotonically increasing final confidence score. This is also a positive correlation. Mcpc reflects the strength of support for the prediction from the "anchor point" in the physical world. The closer a point is to a recent, real physical measurement point, the higher the probability that its prediction will be verified. The "spatial distance with time decay" algorithm of this invention simulates the objective law that the value of physical calibration information decays with time and spatial distance.

[0155] Smst, as a scalar value, plays a global modulating role in the fused confidence map. Its increase or decrease is positively correlated with the final confidence score. Smst is a macroscopic assessment of the system's own "health status." When the prediction residuals for new data points are small, Smst approaches 1, indicating that the model is stable and allowing the confidence assessment from the fused Mmic and Mcpc to output normally. Conversely, when the residuals are large, Smst approaches 0, which globally suppresses the entire confidence distribution map. This reflects that when the model makes a significant error at one point, the confidence of its predictions at all other points should be questioned until the system regains stability with more data.

[0156] To quantitatively verify the beneficial effects of the technical solution of this invention, a high-fidelity digital experimental environment was constructed. This environment simulates a typical scenario in the production of anode carbon blocks where the mapping relationship between the optical and physical properties of the product drifts due to equipment wear. The table below shows the performance comparison between the technical solution of this invention (experimental group) and a static model trained only offline (control group) in three key scenarios. The control group model is not updated after offline training, and its confidence level depends only on the inherent uncertainty of the model. The experiment introduces a novel derived parameter, "confidence error mismatch," to quantify the reliability of the confidence map; see Table 1 below for details.

[0157] Table 1: Performance Comparison of the Invention and Control Group Technical Solutions in Dynamic Production Environments

[0158] Parameter name Scenario 1: Stable Production (Initial) - First Time Scenario 1: Stable Production (Initial) - 2nd Time Scenario 2: Gradual Drift - First Time Scenario 2: Gradual Drift - 2nd Time Scenario 3: Sudden Change in Equipment Status - First Time Scenario 3: Sudden Change in Equipment Status - 2nd Time Model intrinsic confidence (mean) 0.92 0.91 0.88 0.89 0.85 0.86 Calibrate the nearest neighbor confidence level (mean). 0.85 0.86 0.42 0.41 0.35 0.36 Model state credibility 0.98 0.97 0.95 0.96 0.15 0.18 Control group: Actual prediction average error (%) 1.5 1.6 8.2 8.5 15.1 14.8 Control group: Confidence error mismatch 0.05 0.06 0.45 0.48 0.65 0.62 This invention: Actual prediction average error (%) 1.5 1.6 2.1 2 14.5 14.2 This invention: Confidence error mismatch 0.05 0.06 0.08 0.09 0.11 0.12

[0159] The confidence error mismatch, denoted as Smh, is obtained by processing the system output. The calculation steps are as follows: obtain the "estimated confidence distribution map" generated by the system and the "actual prediction error map" obtained by comparing it with the ground truth; linearly normalize all pixel values ​​in both maps to the [0,1] interval; then, calculate the average of the absolute values ​​of the corresponding pixel differences between the normalized confidence map and the "1 minus the normalized error map". This parameter is used to quantitatively characterize the consistency between the system's estimated confidence and its actual prediction accuracy. The lower the value, the more accurate the system's reliability assessment of its prediction results, and the stronger the system's self-awareness. The above experimental data clearly reveal the significant advantages of the technical solution of this invention:

[0160] Adaptive correction capability in gradual drift scenarios: Compared with the data in "Scenario 2", the control group's prediction model gradually failed due to the inability to update online, causing the "actual prediction average error" to soar from the initial 1.5% to 8.2%. More seriously, its "confidence error mismatch" was as high as 0.45, which means that while the system outputs incorrect results, it incorrectly reports a high confidence level, which is misleading.

[0161] Conversely, this invention, through online calibration and model optimization, successfully suppressed the "actual prediction average error" to a low level of 2.1%. Simultaneously, its "confidence error mismatch" was only 0.08. This indicates that the confidence plot of this invention truly reflects the high-precision state after model correction. Compared to the control group, this invention improved prediction accuracy by 74.4%, calculated as (8.2-2.1) / 8.2×100%, and improved the reliability of confidence assessment by 82.2%, calculated as (0.45-0.08) / 0.45×100%.

[0162] Comparing the data with "Scenario 3," which simulates a sudden change in the equipment leading to a large residual between the new calibration point and the model prediction, the model state confidence Smst of this invention drops to 0.15. Although the immediate "actual prediction average error" (14.5%) of this invention is similar to that of the control group (15.1%) due to the sudden change, the two are fundamentally different.

[0163] The "confidence error mismatch" of this invention is only 0.11, while the control group's is as high as 0.65. This demonstrates the core advantage of this invention: when the system encounters an unknown major shock, the model state confidence Smst mechanism reacts immediately, globally lowering the confidence score and issuing a "system state is unreliable" warning to the user. This successfully prevents the system from outputting an incorrect result. The control group, on the other hand, completely lacks this risk perception capability and continues to output misleading high-confidence predictions.

[0164] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0165] The algorithm of this invention is implemented as a Python script. Before executing the core logic, the program first executes a data loading module (e.g., using the widely used pandas library in Python) configured to read the aforementioned spreadsheet file and load its contents into the program's working memory (e.g., a DataFrame data structure). Subsequent algorithm steps will directly query and retrieve the required configuration parameters from this in-memory data structure.

[0166] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A low porosity anode carbon block porosity real time estimation system, characterized by, The specific steps include: An optical data acquisition module is configured to acquire optical micro-morphology data of a surface of the anode carbon block to be measured; A physical property acquisition module is configured to acquire gas transport characteristic data of at least one calibration point of the anode carbon block to be measured; At the calibration point, a tunable diode laser absorption spectroscopy technique is used to measure a transient curve of a concentration of a probe gas changing with time after the probe gas penetrates through the anode carbon block to be measured; and A parameter representing connectivity of a pore network in the anode carbon block to be measured is calculated based on the transient curve, as the gas transport characteristic data; A calibration decision module is configured to generate an initial porosity distribution map by using an online-updatable mapping model based on the optical micro-morphology data, and calculate and output an optimal calibration point instruction according to preset information value criteria of the initial porosity distribution map; A model optimization module is configured to acquire corresponding gas transport characteristic data according to the calibration point determined by the optimal calibration point instruction, and perform incremental optimization of the online-updatable mapping model by using the gas transport characteristic data; and Optical micro-morphology data determined by the optimal calibration point instruction and corresponding gas transport characteristic data form a new training data pair; and A transfer learning or Bayesian updating algorithm is used to fine-tune the mapping model by using the new training data pair. A result generation module is configured to process the optical micro-morphology data by using the mapping model that has been incrementally optimized, to generate a corrected porosity distribution map and an estimated confidence distribution map.

2. A low porosity anode carbon block porosity real time estimation system as claimed in claim 1 wherein: The optical data acquisition module is specifically configured to: A structured light three-dimensional scanning technique is used to project a structured light pattern with a preset code onto the surface of the anode carbon block to be measured; and An image of the structured light pattern modulated by the surface of the anode carbon block to be measured is collected and demodulated, to reconstruct three-dimensional point cloud data of the surface of the anode carbon block to be measured as the optical micro-morphology data.

3. A real time porosity estimation system for low porosity anode blocks as claimed in claim 2 wherein: The calibration decision module is specifically configured to: An information value map representing model inference uncertainty is calculated based on the initial porosity distribution map; and A peak point or a barycenter position in the information value map is selected as a coordinate of the optimal calibration point instruction.

4. A real time porosity estimation system for low porosity anode blocks as claimed in claim 3 wherein: The calibration decision module calculates the information value map, and the steps specifically include: A model uncertainty map is calculated based on the initial porosity distribution map; An optical feature vector of each pixel point is extracted from the optical micro-morphology data, and a feature novelty map is calculated, a value of the feature novelty map representing a deviation degree of the optical feature vector of each pixel point from a historical feature vector set; After the model uncertainty map and the feature novelty map are normalized, they are fused by weighting, to generate a hybrid information value map as the information value map.

5. A real time porosity estimation system for low porosity anode carbon blocks as claimed in claim 4 wherein: When the calibration decision module uses the optical micro-morphology data, it is specifically configured to: Based on three-dimensional point cloud information in the optical micro-morphology data, pore geometric features and spatial distribution features in a local region are identified and quantified; A one-dimensional optical permeability index is calculated by using a pre-calibrated weighting model according to the pore geometric features and the spatial distribution features; and A result generation module is configured to process the optical micro-morphology data by using the mapping model that has been incrementally optimized, to generate a corrected porosity distribution map and an estimated confidence distribution map. The optical data acquisition module is specifically configured to: A structured light three-dimensional scanning technique is used to project a structured light pattern with a preset code onto the surface of the anode carbon block to be measured; and An image of the structured light pattern modulated by the surface of the anode carbon block to be measured is collected and demodulated, to reconstruct three-dimensional point cloud data of the surface of the anode carbon block to be measured as the optical micro-morphology data. The calibration decision module is specifically configured to: An information value map representing model inference uncertainty is calculated based on the initial porosity distribution map; and A peak point or a barycenter position in the information value map is selected as a coordinate of the optimal calibration point instruction. The calibration decision module calculates the information value map, and the steps specifically include: A model uncertainty map is calculated based on the initial porosity distribution map; An optical feature vector of each pixel point is extracted from the optical micro-morphology data, and a feature novelty map is calculated, a value of the feature novelty map representing a deviation degree of the optical feature vector of each pixel point from a historical feature vector set; After the model uncertainty map and the feature novelty map are normalized, they are fused by weighting, to generate a hybrid information value map as the information value map. The calibration decision module uses the optical micro-morphology data, and is specifically configured to: Based on three-dimensional point cloud information in the optical micro-morphology data, pore geometric features and spatial distribution features in a local region are identified and quantified; A one-dimensional optical permeability index is calculated by using a pre-calibrated weighting model according to the pore geometric features and the spatial distribution features; and A result generation module is configured to process the optical micro-morphology data by using the mapping model that has been incrementally optimized, to generate a corrected porosity distribution map and an estimated confidence distribution map. and The optical penetration index is taken as an input of a Gaussian process regression model to generate an initial porosity distribution map and a model uncertainty map.

6. A real time porosity estimation system for low porosity anode blocks as claimed in claim 5 wherein: The result generation module is specifically configured to: Based on the internal uncertainty parameter of the mapping model after incremental optimization, or based on the spatial distance between each point in the initial porosity distribution map and the historical calibration point position, a confidence score corresponding to each point in the corrected porosity distribution map is calculated to form an estimated confidence distribution map.

7. A real time porosity estimation system for low porosity anode blocks as claimed in claim 6 wherein: The step of calculating the confidence score by the result generation module specifically includes: Based on the internal uncertainty parameter of the mapping model, a model internal confidence map is calculated; Based on the spatial distance with time decay between each point and the historical calibration point position, a calibration proximity confidence map is calculated; Based on the residual of the training data pair used in the last incremental optimization of the model, a model state reliability scalar value is calculated; The model internal confidence map and the calibration proximity confidence map are weighted and fused, and the fusion result is modulated using the model state reliability to generate the estimated confidence distribution map.

Images