Bearing steel ball gray cast iron mill plate microstructure performance optimization system based on deep learning

Abstract

The application relates to the technical field of bearing steel ball processing, and discloses a bearing steel ball gray cast iron grinding plate organization performance optimization system based on deep learning, which comprises the following modules: a data acquisition module, which acquires multimodal characteristic data of grinding plate samples, carries out pretreatment and feature processing; an organization feature extraction module, which deeply extracts organization features; a prediction module, which constructs a multimodal quality index prediction model to carry out quality index prediction; a formula optimization module, which carries out collaborative optimization of grinding plate material formula and preparation process; a simulation evaluation module, which simulates the preparation process, records simulation data and simulation results, and carries out feasibility evaluation; a state monitoring module, which carries out grinding plate preparation, monitors preparation process signals in real time, and carries out signal processing; and a service life prediction module, which constructs a quality index degradation prediction model, carries out real-time state evaluation on the grinding plate, and predicts the remaining service life of the grinding plate; the application realizes real-time and accurate evaluation of the working state of the grinding plate.

Description

Technical Field

[0001] This invention relates to the field of bearing steel ball processing technology, and more specifically, to a deep learning-based system for optimizing the microstructure and properties of gray cast iron grinding plates for bearing steel balls. Background Technology

[0002] As a core component of precision machinery, the surface quality and dimensional accuracy of bearing steel balls directly affect the performance and lifespan of the bearing. In the grinding process of steel balls, gray cast iron grinding plates are crucial process equipment; the microstructure and properties of the grinding plates determine the processing quality and production efficiency of the steel balls. However, existing grinding plate designs and manufacturing technologies have many problems, severely restricting the mass production of high-precision bearing steel balls.

[0003] Traditional grinding plate formulation design relies heavily on engineers' experience and trial-and-error methods, lacking scientific theoretical guidance and quantitative design tools. Engineers typically adjust chemical composition ratios based on past production experience, optimizing the formulation through repeated trials and performance tests. This method is not only time-consuming and costly, but also struggles to guarantee the optimality of the formulation. Furthermore, the complex relationship between the microstructure and macroscopic properties of grinding plates is not yet fully understood. The comprehensive influence mechanism of multiple factors, such as graphite morphology, matrix structure, and chemical composition, on grinding plate quality indicators lacks a quantitative descriptive model, resulting in a lack of scientific basis for formulation optimization.

[0004] Furthermore, the degradation patterns of grinding plate quality indicators and the prediction of remaining service life during use have always been challenges in the industry. Traditional maintenance strategies mainly rely on periodic replacement or alarm mechanisms based on simple thresholds, which cannot accurately predict the actual remaining life of the grinding plate. This often results in premature replacement leading to waste or delayed replacement affecting product quality. Existing condition monitoring technologies mostly use single sensor signals, which have limited information dimensions and are easily affected by noise interference, making it difficult to comprehensively reflect the true working condition of the grinding plate. The lack of effective multimodal data fusion and intelligent analysis methods makes it impossible to extract valuable condition features from complex industrial data, resulting in low accuracy in condition assessment and life prediction. Summary of the Invention

[0005] This invention provides a deep learning-based system for optimizing the microstructure and properties of gray cast iron grinding plates for bearing steel balls, solving the technical problems in related technologies such as grinding plate formulation design relying on experience, inaccurate prediction of quality indicators, and difficulty in predicting service life.

[0006] This invention provides a deep learning-based system for optimizing the microstructure and properties of gray cast iron grinding plates for bearing steel balls, comprising: The data acquisition module collects multimodal feature data of the grinding plate sample, performs preprocessing and feature processing, and obtains a multimodal fusion sample dataset. The tissue feature extraction module, based on a multimodal fusion sample dataset, deeply extracts tissue features to obtain a micro-tissue feature vector dataset; The prediction module, based on the micro-organism feature vector dataset and the multimodal fusion sample dataset, constructs a multimodal quality indicator prediction model to predict quality indicators and obtain the quality indicator prediction results. The formulation optimization module, based on the quality index prediction results, performs synergistic optimization of the grinding plate material formulation and preparation process to obtain the grinding plate preparation scheme; The simulation evaluation module, based on the grinding plate preparation scheme, simulates the preparation process and records the simulation data and results to conduct a feasibility evaluation and obtain the scheme feasibility evaluation results. The state monitoring module, based on the feasibility assessment results of the scheme, carries out grinding plate preparation according to the grinding plate preparation scheme, monitors the preparation process signals in real time and performs signal processing to obtain the comprehensive state feature vector of the grinding plate; The life prediction module constructs a quality index degradation prediction model, combines the comprehensive state feature vector of the grinding plate to perform real-time state assessment of the grinding plate, predicts the remaining service life of the grinding plate, and obtains the remaining service life prediction results and maintenance decision suggestions.

[0007] In a preferred embodiment, the data acquisition module includes acquiring microstructure image data of the grinding plate sample, processing the grinding plate sample using a standard metallographic sample preparation process, cutting a sample block from a preset depth region below the working surface of the grinding plate, performing coarse grinding, fine grinding and polishing treatments in sequence, performing etching treatment with a nitric acid alcohol solution of a preset concentration, and observing and photographing the sample at multiple magnifications using a metallographic microscope to obtain a raw database of metallographic structure images.

[0008] In a preferred embodiment, the data acquisition module further includes: Chemical composition data of the grinding plate sample was obtained, and the chemical composition was quantitatively detected by a spectrometer to obtain the mass fraction data of the elements in the grinding plate. Sample blocks suitable for spectrometer analysis were cut from the grinding plate, and the sample surface was polished to remove the oxide layer and contaminants. The sample was excited using a direct-reading spectrometer, and the intensity of characteristic spectral lines was measured using the direct-reading spectrometer. The intensity of the characteristic spectral lines was converted into elemental content data. For each sample, the measurement was repeated at different positions for no less than a preset number of times, and the average value of the measurement results was taken as the chemical composition data of the sample.

[0009] In a preferred embodiment, the tissue feature extraction module includes: A metallographic feature extraction model was constructed by loading the network structure and weight parameters of a pre-trained convolutional neural network model, retaining the convolutional layer part of the pre-trained convolutional neural network model as a feature extractor, and removing the fully connected classification layer of the original model. A new layer structure is added to the top of the pre-trained convolutional neural network model, and a hierarchical freezing strategy is used for training. In the early stage of training, the parameters of the bottom convolutional layers of the pre-trained model are frozen, and only the newly added top layer structure is trained. In the later stage of training, some high-level convolutional layers are unfrozen, and some high-level convolutional layers are fine-tuned.

[0010] In a preferred embodiment, the tissue feature extraction module further includes: Design a deep convolutional neural network model for extracting tissue features. The model adopts a two-branch convolutional neural network structure. The first branch is used for the identification and quantification of graphite features, and the second branch is used for the identification and quantification of matrix tissue features. Within each branch, an attention mechanism module is introduced, which includes two types: spatial attention and channel attention. The spatial attention mechanism assigns different weights to different positions based on the spatial importance of the feature map, while the channel attention mechanism assigns different weights to different channels based on the importance of the feature channels. The feature vectors output by the two branches are then concatenated or weighted and fused.

[0011] In a preferred embodiment, the prediction module includes: Construct an initial quality indicator prediction model and design a multi-input multi-output fully connected deep neural network architecture, which includes two independent input branches; The first input branch receives the microscopic tissue feature vector, which consists of multiple fully connected layers. Each layer contains several neuron nodes, and the nodes are fully connected. Non-linearity is introduced by an activation function. The second input branch receives the chemical composition feature vector and consists of multiple fully connected layers. After each of the two input branches completes several layers of transformation, feature fusion is performed in the middle layer of the fully connected deep neural network.

[0012] In a preferred embodiment, the prediction module further includes: The Bayesianization of the initial quality index prediction model transforms the weight parameters in the initial quality index prediction model from fixed values ​​into random variables that follow a probability distribution. Each weight is described by a probability distribution with mean and variance, assuming that the prior distribution of the weights is a standard normal distribution. An approximate inference is performed using variational inference. It is assumed that the posterior distribution follows a diagonal Gaussian distribution. Each weight is described by two parameters: mean and variance. The reparameterization technique is used to make the gradient backpropagate through random sampling. The loss function of variational inference is defined.

[0013] In a preferred embodiment, the formula optimization module includes: The candidate optimization solution set for the chemical composition of the grinding plate is obtained based on the genetic algorithm. The optimization variable is defined as the chemical composition parameter of the grinding plate. The upper and lower bounds of the value of each element are set according to the composition range of the grinding plate material. The genetic algorithm population is initialized by randomly generating a preset number of individuals. Each individual represents a set of chemical composition formulas and is encoded as a real number vector. Each component of the vector corresponds to the quality score of each element. The fitness of each individual in the population is evaluated, and the Pareto dominance relation is used to evaluate the quality of individuals. An evolutionary cycle of selection, crossover, mutation, and elite retention is performed based on fitness ranking and crowding distance.

[0014] In a preferred embodiment, the simulation evaluation module includes: A digital twin simulation model of the grinding process is obtained based on multiphysics modeling; Establish geometric models for the grinding plate and the steel ball. The grinding plate is modeled as a disk-shaped structure with preset curvature and surface texture, and the steel ball is modeled as a standard sphere. Establish material models to describe the mechanical properties of the grinding plate and the steel ball. A contact mechanics model is established to describe the contact behavior between the grinding plate and the steel ball, and the contact stress distribution is calculated; a wear model is established to describe the material removal mechanism during the grinding process; a friction model is established to describe the friction behavior of the contact interface; a heat conduction model is established to describe the temperature field evolution during the grinding process; and a surface morphology evolution model is established to describe the dynamic changes in the surface roughness and sphericity of the steel ball. The geometric model, material model, contact mechanics model, wear model, friction model, heat conduction model, and surface morphology evolution model are coupled.

[0015] In a preferred embodiment, the lifetime prediction module includes: A quality indicator degradation prediction model was built and trained based on historical data; Based on the quality index degradation prediction model and the comprehensive state feature vector of the grinding plate, the current state of the grinding plate is evaluated in real time, and the real-time evaluation result of the current state of the grinding plate is obtained. Based on the real-time evaluation results of the current state of the grinding plate and the quality index degradation prediction model, the trend extrapolation method is used to predict the future quality index degradation trajectory of the grinding plate, and obtain the remaining service life prediction results and maintenance decision suggestions.

[0016] The beneficial effects of this invention are as follows: This invention effectively solves key technical problems in the design and use of traditional grinding plates by constructing a deep learning-based multimodal data fusion and intelligent optimization system. The system employs multimodal feature extraction and fusion technology to deeply integrate metallographic images, chemical composition data, and physical performance test data, establishing a quantitative mapping relationship between microstructure and macroscopic quality indicators. A Bayesian neural network is used to quantitatively predict the uncertainty of quality indicators, providing a reliable theoretical basis for formulation optimization. The synergistic application of genetic algorithms and Bayesian optimization achieves global optimization of chemical composition and heat treatment processes, improving the scientific rigor and efficiency of grinding plate formulation design, shortening product development cycles, and reducing trial production costs.

[0017] This invention establishes a multi-sensor condition monitoring and lifespan prediction mechanism. Through adaptive Kalman filtering and multi-sensor data fusion technology, it achieves real-time and accurate assessment of the grinding plate's working status. Based on a quality index degradation prediction model using a long short-term memory network, combined with an online learning mechanism, it can accurately predict the remaining service life of the grinding plate and provide maintenance decision suggestions. The application of digital twin simulation technology allows the optimized scheme to be fully validated before practical application, reducing implementation risks. The entire system forms a complete closed loop from formula design and manufacturing process optimization to usage status monitoring, realizing intelligent management of the entire life cycle of the grinding plate, effectively improving the processing quality and production efficiency of bearing steel balls, and providing important technical support for high-end bearing manufacturing. Attached Figure Description

[0018] Figure 1 This is a block diagram of the deep learning-based microstructure and performance optimization system for gray cast iron grinding plates for bearing steel balls in this invention. Figure 2 This is a detailed flowchart of the deep learning-based microstructure and performance optimization system for gray cast iron grinding plates for bearing steel balls in this invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a deep learning-based system for optimizing the microstructure and properties of gray cast iron grinding plates for bearing steel balls, such as... Figures 1 to 2 As shown, it includes: The data acquisition module collects multimodal feature data of the grinding plate sample, performs preprocessing and feature processing, and obtains a multimodal fusion sample dataset. Step 1.1: Obtain microstructure image data of the grinding plate sample; Based on the grinding plate sample to be optimized, a standard metallographic sample preparation procedure was adopted to obtain a metallographic specimen that meets the requirements for microscopic observation. Specifically, a representative sample block was cut from the grinding plate, selecting a region 3 to 5 mm below the working surface of the grinding plate, as the microstructure in this region best reflects the material properties of the grinding plate during the actual grinding process. The cut specimens were subjected to coarse grinding, fine grinding, and polishing in sequence. Different grit sandpapers were used to grind the specimen surface from coarse to fine to remove the deformation layer and scratches generated during the cutting process. Finally, diamond polishing paste was used to polish the surface to obtain a mirror-like observation surface. The polished specimen surface was etched using a 4% (v / v) nitric acid alcohol solution. The etching time was adjusted according to the carbon content and microstructure type of the specimen. The etching process created a high degree of difference in the microstructure of different phases, thus revealing a clear microstructure morphology under the microscope. Metallurgical microscopes were used to observe and photograph the corroded samples at multiple magnifications. Images were taken at 100x magnification to observe the overall distribution of graphite, and at 500x magnification to observe the relative content and morphological characteristics of pearlite and ferrite in the matrix. For each sample, at least 20 images were taken from different fields of view to ensure that the acquired images represent the microstructure distribution of the entire sample. The original metallurgical images were digitally stored, recording the shooting parameters for each image, including magnification, exposure time, and light source intensity, resulting in a fully annotated original database of metallurgical images.

[0021] Step 1.2: Obtain the chemical composition data of the grinding plate sample; Based on the grinding plate sample to be optimized, a spectrometer was used to quantitatively detect its chemical composition, obtaining the mass fraction data of each major element and alloying element in the grinding plate. Specifically, a sample block suitable for spectrometer analysis was cut from the grinding plate, and the sample surface was polished to remove the oxide layer and contaminants, ensuring a clean and flat surface for testing. A direct-reading spectrometer was used to excite the sample; the high-energy arc or spark of the spectrometer caused the material on the sample surface to evaporate and generate characteristic spectra. The intensity of the characteristic spectral lines was measured by the spectrometer's dispersive system and detector, and the spectral line intensity was converted into elemental content data according to a pre-established working curve. The elements detected included major elements such as carbon, silicon, manganese, phosphorus, and sulfur, as well as potentially present alloying elements and trace elements such as nickel, chromium, molybdenum, vanadium, titanium, aluminum, and copper. For each sample, at least three repeated measurements were performed at different locations, and the average value of the measurement results was taken as the chemical composition data of that sample. The standard deviation was calculated to assess the compositional uniformity. For particularly critical elements such as carbon and silicon, combustion methods or chemical analysis were used for verification to ensure the accuracy of the composition data. The obtained chemical composition data is organized into a structured data table, with each row corresponding to a grinding plate sample and each column corresponding to the mass fraction of a chemical element, thus obtaining the original database of chemical composition of the grinding plate.

[0022] Step 1.3: Obtain the physical property test data of the grinding plate sample; Based on the grinding plate sample to be optimized, standard physical performance testing methods were used to obtain key quality index data such as hardness and wear resistance, resulting in a dataset characterizing the macroscopic performance of the grinding plate. Specifically, a Rockwell hardness tester was used to test the hardness of the grinding plate. Appropriate indenters and loads were selected, and multiple measurements were taken on the working surface of the grinding plate. The measurement points were evenly distributed and spaced according to standard requirements to avoid mutual interference between adjacent measurement points. The hardness value of each measurement point was recorded, and the average hardness and standard deviation of the hardness distribution were calculated. The average hardness reflects the overall hardness level of the grinding plate, and the standard deviation reflects the uniformity of hardness. A wear resistance tester was used to test the wear resistance of the grinding plate. The grinding plate sample was subjected to friction and wear with a standard abrasive under specified load and speed conditions. The mass loss or volume loss during the wear process was recorded, and the wear rate per unit time or unit stroke was calculated. The magnitude of the wear rate directly reflects the wear resistance performance of the grinding plate. A roughness meter was used to measure the surface roughness of the working surface of the grinding plate. Roughness parameters included the arithmetic mean deviation of the profile and the maximum profile height. Surface roughness affects the contact state between the grinding plate and the steel ball and the grinding efficiency. By associating test data such as hardness, wear resistance, and surface roughness with the corresponding grinding plate sample numbers, a database of grinding plate physical properties is established.

[0023] This step completes the data acquisition, yielding multimodal feature data including a fully annotated original database of metallographic images, an original database of the chemical composition of grinding plates, and a database of the physical properties of grinding plates. Step 1.4: Obtain a standardized metallographic structure image dataset based on image preprocessing techniques; Based on the original metallographic image database obtained in step 1.1, digital image processing techniques were used to preprocess and standardize the original images, resulting in a high-quality image dataset suitable for deep learning model input. Specifically, the original images underwent format conversion and size normalization, converting images acquired from different shooting devices or at different times to the same image format and color space, and adjusting the image size to a uniform pixel resolution. Median filtering or Gaussian filtering was used to denoise the images, removing random noise and salt-and-pepper noise introduced during image acquisition while maintaining the clarity of tissue boundaries. Histogram equalization or adaptive contrast enhancement techniques were used to adjust the contrast of the images, enhancing the grayscale differences between different tissue phases in the image, making the boundaries between graphite and matrix, and pearlite and ferrite clearer. Geometric correction was performed on the images to compensate for image distortion caused by the microscope optical system or shooting angle. For images with uneven illumination, a background correction algorithm was used to eliminate the gradient phenomenon of bright center and dark periphery caused by bright field illumination. The processed images are labeled and stored according to a unified naming rule. The image file name includes information such as sample number, magnification, and field of view position, which facilitates subsequent data management and traceability, resulting in a standardized metallographic structure image dataset.

[0024] Step 1.5: Obtain enhanced chemical composition feature vectors based on feature engineering methods; Based on the original chemical composition database obtained in step 1.2, feature engineering was performed using expertise in materials science to obtain more expressive chemical composition feature vectors. Specifically, based on the original elemental mass fraction features, derived features were calculated to reflect the synergistic effects between elements and their comprehensive influence on microstructure and properties. The carbon equivalent value was calculated. Carbon equivalent is an important parameter characterizing the degree of graphitization in cast iron. The formula is: carbon equivalent = carbon mass fraction + silicon mass fraction divided by 4 + phosphorus mass fraction divided by 2. The value of carbon equivalent directly affects the precipitation behavior of graphite and the matrix microstructure type. The eutectic degree was calculated. Eutectic degree represents the degree to which the cast iron composition approaches the eutectic point. It is determined by the ratio of carbon equivalent to eutectic carbon equivalent. Eutectic degree affects the solidification mode and final microstructure of cast iron. The silicon-carbon ratio was calculated. The silicon-carbon ratio reflects the content relationship between silicon and carbon. The magnitude of the silicon-carbon ratio affects the morphology of graphite and the proportion of ferrite in the matrix. The alloying coefficient is calculated. For cast iron with added alloying elements such as nickel, chromium, and molybdenum, a comprehensive alloying coefficient is calculated based on the mass fraction of each alloying element and its weight in relation to microstructure and properties. For the mass fraction of each element, its deviation relative to the standard formulation or industry average is calculated; this deviation characteristic reflects the specificity of the current sample composition. The chemical composition characteristics are normalized using minimum-maximum normalization or standardization methods to map the mass fractions of different elements to the same numerical range, eliminating the influence of differences in element content magnitudes and ensuring a relatively balanced contribution of each feature to the model. The original element mass fractions, the calculated derived features, and the normalized features are combined to form a comprehensive chemical composition feature vector. This feature vector serves as one of the inputs to the subsequent model, resulting in a chemical composition feature vector dataset.

[0025] In some embodiments, due to differences in raw material sources and formulation systems used by different manufacturers, standard feature engineering methods may not be able to fully capture the unique component-performance relationships of each manufacturer. Therefore, methods based on automatic feature learning can be used to extract deep features from chemical component data. The aim is to uncover the nonlinear relationships and higher-order interactions hidden in the original component data, thereby improving the model's adaptability to different formulation systems. Specifically, an autoencoder neural network is constructed, consisting of an encoder and a decoder. The encoder consists of multiple fully connected layers that compress the original features of the input chemical components layer by layer into low-dimensional latent feature representations. The decoder restores the latent features to the original feature space, and the network is trained by minimizing the difference between the input and the reconstructed output. After training, the latent features output by the encoder are extracted as deep features of the chemical components. These deep features are abstract representations of the original component data after nonlinear transformation, and can more effectively reflect the potential impact of components on performance. The automatically learned deep features are concatenated with manually designed engineered features to form a hybrid feature vector, combining the advantages of domain knowledge and data-driven approaches to further enhance the expressive power of the features.

[0026] Step 1.6: Obtain a multimodal fusion sample dataset based on data association and synchronization; Based on the standardized metallographic image dataset obtained in step 1.4, the chemical composition feature vector dataset obtained in step 1.5, and the physical property database obtained in step 1.3, data association and synchronization are performed using the sample number as the association key to obtain a complete multimodal data record for each sample. Specifically, using the unique number of the grinding plate sample as an index, all relevant data for the corresponding sample are extracted from various data sources, and the metallographic image file path, chemical composition feature vector, and physical property test values ​​of the same sample are organized into a complete data record. For samples with multiple metallographic images, a correspondence between the image sequence and the sample number is established to ensure that all images of a sample can be correctly associated. The integrity of the data record is checked, and samples missing certain modal data are identified. For missing values, an appropriate handling strategy is adopted. If the missing proportion is small and the sample is important, the missing data is obtained through supplementary testing; if the missing proportion is large or the sample is not representative, the sample is removed from the dataset to ensure that each sample in the final dataset contains complete multimodal information. The associated dataset is divided into training, validation, and test sets according to a certain ratio. The training set is used for learning model parameters, the validation set is used for adjusting model hyperparameters and implementing early stopping strategies, and the test set is used for evaluating the final model performance. During the partitioning process, it is ensured that the samples in different subsets have similar distributions in chemical composition range, tissue type, and performance level to avoid biases introduced by data partitioning affecting the model evaluation results. The information of the partitioned dataset is recorded in a metadata file, including the sample list, total number of samples, and feature dimensions of each subset, resulting in a multimodal fusion sample dataset.

[0027] The tissue feature extraction module, based on a multimodal fusion sample dataset, deeply extracts tissue features to obtain a micro-tissue feature vector dataset; Step 2.1: Construct a metallographic structure feature extraction model; Based on pre-trained convolutional neural network models widely used in the field of deep learning, a transfer learning method is employed as the starting point for feature extraction from the standardized metallographic image dataset obtained in step 1.4. Specifically, classic convolutional neural network architectures pre-trained on large-scale natural image datasets such as ImageNet are selected, including VGG, ResNet, or EfficientNet.

[0028] While standardized metallographic image datasets differ somewhat in content from natural images, they share common underlying geometric and textural features. Utilizing a pre-trained convolutional neural network (CNN) model can reduce the need for labeled data. The network structure and weight parameters of the pre-trained CNN model are loaded, retaining the convolutional layers as feature extractors. Since the original classification layer is designed for natural image categories and is unsuitable for metallographic analysis, its fully connected classification layer is removed. A new layer structure adapted for metallographic analysis is added to the top of the pre-trained CNN model, including a global average pooling layer, fully connected layers, and an output layer. The design of the output layer is determined based on the specific task objective; for the tissue feature extraction task, the output layer outputs a fixed-dimensional feature vector. A layer-freezing strategy is employed for model training. Initially, the parameters of the lower convolutional layers are frozen, and only the newly added top layer structure is trained, allowing the top layer to quickly adapt to the feature distribution of the metallographic images. Later in training, some higher-level convolutional layers are unfrozen, and these layers are fine-tuned to enable the pre-trained CNN model to learn representations that better reflect the metallographic features. Through transfer learning, the model can achieve good feature extraction performance on a relatively small labeled dataset, resulting in a metallographic feature extraction model.

[0029] Step 2.2, expand the image dataset; Based on the standardized metallographic image dataset obtained in step 1.4, new training image samples are generated using data augmentation techniques to obtain an expanded training image dataset. Specifically, geometric transformations are performed on the original images, including random rotation, horizontal flip, vertical flip, translation, scaling, and cropping. Random rotation rotates the image around its center point by a random angle, ranging from -90 degrees to +90 degrees; the rotated image is then filled using interpolation to maintain its original size. Horizontal and vertical flips mirror the image along the horizontal or vertical axis, reflecting the statistical symmetry of metallographic structures in different directions. Translation moves the image a certain pixel distance horizontally or vertically; the resulting blank areas are filled with boundary pixel values ​​or mirror images. Scaling enlarges or reduces the image to simulate different viewing magnifications; after scaling, the image is restored to its standard size through cropping or filling. Cropping applies non-uniform geometric deformation to the image to simulate small deviations in the shooting angle.

[0030] Pixel-level transformations are performed on the original image, including brightness adjustment, contrast adjustment, saturation adjustment, and noise addition. Brightness adjustment alters the overall brightness by adding or subtracting pixel values, simulating changes in the intensity of a microscope light source. Contrast adjustment changes the image's contrast by expanding or compressing its grayscale dynamic range, simulating different exposure conditions. Noise addition introduces Gaussian or salt-and-pepper noise into the image, simulating electronic noise and environmental interference during image acquisition, thus improving the model's robustness to noise.

[0031] For each original image, multiple transformed images are generated by randomly combining the above transformation operations. The transformation parameters are randomly selected within a reasonable range to ensure that the generated images are visually plausible and undistorted. The generated enhanced images are added to the training dataset and used together with the original images for model training. The size of the augmented training set is increased several times, resulting in an expanded training image dataset.

[0032] Step 2.3: Design a deep convolutional neural network model for tissue feature extraction; Based on the metallographic feature extraction model obtained in step 2.1 and the expanded training image dataset obtained in step 2.2, a multi-branch network architecture is designed to extract specific features of the metallographic structure, resulting in a deep convolutional neural network model for feature extraction. Specifically, a two-branch convolutional neural network structure is designed, with both branches sharing the bottom-level feature extraction convolutional layer and branching into two independent processing paths in the middle layer of the network. The first branch is dedicated to the identification and quantification of graphite features. The network layer of this branch is optimized for the morphological features of graphite. Through the design of the convolutional kernel size and receptive field, the network can effectively capture the morphological features of graphite, such as its elongated, flake, or spherical shapes, as well as parameters such as graphite size, aspect ratio, and distribution density. The second branch is dedicated to the identification and quantification of matrix structure features. This branch is optimized for the texture features of pearlite and ferrite, and can identify the interlamellar spacing of pearlite, the grain size of ferrite, and the relative content of the two phases. Within each branch, an attention mechanism module is introduced, including two types of attention mechanisms: spatial attention and channel attention.

[0033] The spatial attention mechanism assigns different weights to different locations based on the spatial importance of the feature maps, enabling the network to automatically focus on image regions containing key tissue features and suppress interference from background and irrelevant areas. The channel attention mechanism assigns different weights to different channels based on the importance of the feature channels, allowing the network to strengthen feature dimensions that significantly impact performance and weaken redundant feature channels. The feature vectors output from the two branches are concatenated or weighted and fused to form a comprehensive metallographic feature representation. This feature vector simultaneously contains information on graphite features and matrix tissue features, comprehensively describing the microstructure of the grinding plate. A multi-task learning framework is used to train the network, with multiple related supervised tasks designed, including graphite type classification, graphite morphology parameter regression, matrix tissue type classification, and phase content regression. Different tasks share the underlying feature extraction network, improving the overall feature learning effect through knowledge sharing and complementarity between tasks. A comprehensive loss function is defined, and the loss functions of each task are weighted and summed. Backpropagation is used to simultaneously optimize all tasks, with the weights adjusted according to the importance of each task to the final quality index prediction. The network is trained using the augmented training image dataset obtained in step 2.2. Stochastic gradient descent or its variants are used to update the network parameters, with appropriate learning rates, batch sizes, and training epochs set. During training, a validation set is used to monitor model performance. Training is stopped when the loss on the validation set no longer decreases to avoid overfitting. After training, a deep convolutional neural network model for tissue feature extraction is obtained.

[0034] Step 2.4: Obtain the quantitative microstructure feature vector of the grinding plate sample based on the trained model; Based on the deep convolutional neural network model for extracting microstructure features obtained in step 2.3 and the standardized metallographic image dataset, a model inference method is used to extract the deep feature representation of the images, resulting in quantitative microstructure feature vectors. Specifically, the standardized metallographic image of the grinding plate to be analyzed is input into the trained convolutional neural network model. The image undergoes convolution, activation, pooling, and other operations at each layer of the model, extracting and abstracting features layer by layer. During the forward propagation of the network, the bottom convolutional layers extract local edge and texture features of the image, the middle convolutional layers combine local features to form more complex patterns, and the high-level convolutional layers extract abstract features with semantic meaning. Graphite-related feature vectors are extracted from the output layer of the graphite feature branch of the network. Each dimension of this vector corresponds to different feature parameters of graphite, such as the probability distribution of graphite type, average length, average width, aspect ratio, area fraction, and distribution uniformity. Matrix-related feature vectors are extracted from the output layer of the matrix microstructure feature branch of the network. These vectors reflect parameters such as the relative content of pearlite and ferrite, the interlamellar spacing of pearlite, and the grain size of ferrite. The feature vectors from the two branches are concatenated to form a comprehensive microstructure feature vector. This vector is a quantitative numerical representation of the metallographic image of the polished plate, and its dimension is typically between 100 and 500, with the specific dimension determined by the network architecture design. For multiple metallographic images of the same polished plate sample, the feature vectors of each image are extracted, and the average of all feature vectors is calculated as the representative microstructure feature of the sample. The averaging operation reduces the randomness of individual images and provides a more stable description of the microstructure features. The extracted microstructure feature vectors are associated with sample numbers and stored to form a sample-microstructure feature database. Each sample in this database corresponds to a feature vector of a fixed dimension, resulting in a dataset of microstructure feature vectors for the polished plate samples.

[0035] The prediction module, based on the micro-organism feature vector dataset and the multimodal fusion sample dataset, constructs a multimodal quality indicator prediction model to predict quality indicators and obtain the quality indicator prediction results. Step 3.1: Construct an initial quality indicator prediction model; Based on the microstructure feature vector dataset obtained by the tissue feature extraction module and the chemical composition feature vector dataset obtained by the data acquisition module, a multimodal feature fusion quality index prediction model is constructed using a deep neural network to obtain the initial quality index prediction model.

[0036] Specifically, a multi-input multi-output (MIMO) fully connected deep neural network architecture is designed, comprising two independent input branches. The first input branch receives microscopic tissue feature vectors and consists of multiple fully connected layers, each containing several neurons. These neurons are fully connected, and ReLU or LeakyReLU is used as the activation function to introduce non-linearity. The first input branch abstracts and transforms the tissue features layer by layer, learning the complex mapping relationship between tissue features and quality indicators. The second input branch receives chemical composition feature vectors and consists of multiple fully connected layers. It transforms and encodes the component features, learning the correlation between components and quality indicators. After each input branch completes several layers of transformation, feature fusion is performed in the intermediate layer of the fully connected deep neural network. Fusion methods include simple concatenation fusion and attention-weighted fusion. Simple concatenation fusion directly concatenates the feature vectors output from the two branches to form a higher-dimensional joint feature vector. Attention-weighted fusion learns an attention weight vector and dynamically adjusts the relative importance of tissue and component features based on the characteristics of the current sample, weighting the features from the two branches.

[0037] The fused joint feature vector is input into a shared deep network, which further extracts the interaction and synergy of the two modalities, learning a higher-level abstract representation. At the network's output, multiple parallel output heads are designed, each responsible for predicting a quality indicator. These output heads include prediction heads for grinding plate hardness, wear resistance, grinding efficiency, influence on steel ball surface roughness, and influence on steel ball sphericity accuracy. Each output head consists of one or more fully connected layers, with the number of neurons in the last layer corresponding to the dimension of the quality indicator. For continuous numerical indicators, a single numerical value is output; for categorical indicators, a category probability distribution is output. A multi-task learning framework is used to train all output heads simultaneously. The comprehensive loss function is defined as the weighted sum of the loss functions for each task. Mean squared error loss is used for regression tasks, and cross-entropy loss is used for classification tasks. By sharing the underlying feature representation, different quality indicator prediction tasks can mutually promote each other, leveraging the correlation between tasks to improve overall prediction accuracy. The network is trained using the training dataset. The network weights are updated using either the Adam or SGD optimization algorithm. An appropriate learning rate scheduling strategy is set: a larger learning rate is used in the early stages of training for rapid convergence, and a smaller learning rate is used in the later stages for fine-tuning. The model performance is evaluated using the validation dataset. Based on the prediction error on the validation set, the model hyperparameters, including the number of network layers, the number of neurons per layer, the learning rate, and the regularization coefficient, are adjusted. Training stops when the performance on the validation set no longer improves, yielding the initial quality metric prediction model.

[0038] Step 3.2, Bayesianize the initial quality index prediction model; Based on the initial quality index prediction model obtained in step 3.1, Bayesian deep learning technology is used to transform the deterministic model into a Bayesian neural network, resulting in a multimodal quality index prediction model. Specifically, the weight parameters in the initial quality index prediction model are transformed from fixed values ​​into random variables following a certain probability distribution. Each weight is no longer a fixed value but a probability distribution described by the mean and variance. It is assumed that the prior distribution of the weights is a standard normal distribution or other suitable probability distribution, expressing the initial belief about the weight parameters before training. Using Bayesian inference, the prior distribution of the weights is updated using training data to obtain the posterior distribution. The posterior distribution integrates prior knowledge and information from observed data, reflecting the possible values ​​of the weights and their uncertainties under given training data conditions. Due to the high dimensionality and nonlinearity of neural networks, accurate inference of the posterior distribution of weights is computationally infeasible; therefore, variational inference is used for approximate inference. Variational inference transforms the posterior inference problem into an optimization problem, assuming that the posterior distribution belongs to a tractable family of distributions, such as a diagonal Gaussian distribution. That is, it is assumed that each weight is independent and follows a normal distribution, and each weight is described by two parameters: mean and variance. The KL divergence between the variational distribution and the true posterior distribution is defined as the optimization objective. Minimizing the KL divergence makes the variational distribution as close to the true posterior distribution as possible. A reparameterization technique is employed to allow backpropagation of the gradient through random sampling. Specifically, the weights are represented as random samples of a deterministic mean plus standard deviation multiplied by a standard normal distribution. This randomness of the weights is introduced through an independent noise source, allowing the gradient to be calculated using the mean and standard deviation parameters. A loss function for variational inference is defined, consisting of two parts: the first is the data fitting term, i.e., the prediction error; the second is the KL divergence regularization term, which measures the difference between the variational distribution and the prior distribution. The variational parameters, i.e., the mean and variance of the weight distribution, are optimized using the training dataset. The parameters are updated using stochastic gradient descent. In each iteration, a set of specific weight values ​​is sampled from the weight distribution. These weights are used for forward propagation and loss calculation, and then the distribution parameters are updated through backpropagation. After training convergence, the posterior distribution parameters of all weights are output, resulting in a multimodal quality index prediction model.

[0039] Step 3.3, predict quality indicators; Based on the multimodal quality index prediction model, microstructure feature vector dataset, and chemical composition feature vector obtained in step 3.2, a Monte Carlo sampling method is used to generate the probability distribution of the prediction results, yielding a quality index prediction output containing the prediction mean and uncertainty interval. Specifically, for a grinding plate sample to be predicted, its microstructure feature vector and chemical composition feature vector are input into a Bayesian neural network model. Since the network weights are probability distributions rather than deterministic values, each forward propagation requires sampling a specific set of weight parameters from the posterior distribution of the weights. The sampled weights are then used to calculate a prediction output. This sampling and forward propagation process is repeated multiple times, sampling different weight combinations each time, resulting in multiple prediction output values. The set of these output values ​​constitutes the empirical distribution of the prediction results. Typically, 100 to 1000 samplings are performed. The more samplings, the more accurate the estimation of the prediction distribution, but the higher the computational cost. Statistical analysis is performed on the multiple predicted values ​​obtained from the sampling, and the prediction mean is calculated as the final prediction result. The prediction mean is the arithmetic mean of all sampled prediction values, representing the expected value of the prediction. The standard deviation or variance of the predictions is calculated as a measure of prediction uncertainty. A larger standard deviation indicates higher prediction uncertainty and lower model confidence in the sample. Confidence intervals (e.g., 95% confidence intervals) are calculated for the predicted values. The sampled predicted values ​​are sorted by size, and the 2.5% and 97.5% quantiles are used as the lower and upper bounds of the confidence interval. The width of the confidence interval directly reflects the prediction uncertainty. The sources of prediction uncertainty are distinguished, and the total uncertainty is decomposed into cognitive uncertainty and random uncertainty. Cognitive uncertainty stems from the uncertainty of model parameters, reflecting a lack of knowledge due to insufficient training data or model structural limitations. This can be reduced by increasing training data or improving the model structure. Random uncertainty originates from random noise in the data itself and is an inherent uncertainty that cannot be eliminated. By analyzing the variation patterns of predicted values ​​in different samples, the relative contributions of the two types of uncertainty can be estimated. The prediction mean, prediction standard deviation, confidence intervals, and uncertainty decomposition results are organized into a structured prediction output. For each quality indicator, the output format includes the positive or negative uncertainty of the predicted value, as well as the range of confidence levels, resulting in the quality indicator prediction results, which include the prediction mean and uncertainty.

[0040] The formulation optimization module, based on the quality index prediction results, performs synergistic optimization of the grinding plate material formulation and preparation process to obtain the grinding plate preparation scheme; Step 4.1: Obtain the multi-objective optimization function for the grinding plate quality index based on the processing requirements analysis; Based on the specifications and processing quality standards of the target bearing steel balls, an engineering analysis method is used to determine the performance targets that the grinding plate should achieve, resulting in an objective function system containing multiple optimization objectives and constraints. Specifically, according to the application field and precision level of the steel balls, key quality indicators such as surface roughness requirements, sphericity accuracy requirements, and dimensional consistency requirements of the finished steel balls are clarified. For steel balls used in CNC machine tool spindle bearings, the surface roughness requirement is to achieve a profile arithmetic mean deviation within 0.012 micrometers, the sphericity accuracy requirement is to achieve a deviation within 0.5 micrometers, and the dimensional dispersion requirement is to control it within 1 micrometer. The quality requirements of the steel balls are translated into requirements for the quality indicators of the grinding plate, analyzing which performance parameters of the grinding plate directly affect the quality of the steel balls. The hardness of the grinding plate affects the grinding efficiency and grinding plate life. Too low a hardness leads to rapid wear of the grinding plate and a decrease in grinding efficiency, while too high a hardness may cause scratches on the surface of the steel balls. The wear resistance of the grinding plate determines its service life and performance stability; a grinding plate with good wear resistance can maintain stable grinding performance for a long time. The uniformity of the microstructure on the surface of the grinding plate affects the uniformity of the steel ball processing. Inhomogeneous microstructure can lead to ripples or excessive roughness on the surface of the steel ball.

[0041] A multi-objective optimization function is constructed. The first objective is to maximize the service life of the grinding plate, which is directly proportional to wear resistance. Improving wear resistance extends the effective working time of the grinding plate, reducing replacement frequency and production interruptions. The second objective is to maximize grinding efficiency, defined as the amount of material removed from the surface of the steel balls per unit time or the number of steel balls achieving the target surface quality per unit time. Grinding efficiency is related to the hardness and microstructure of the grinding plate. The third objective is to minimize the adverse effects on the surface quality of the steel balls. A predictive model estimates the impact of the grinding plate on the surface roughness and sphericity accuracy of the steel balls, aiming to minimize these influencing indicators and ensure that the processed steel balls meet quality standards. There are interrelationships among the various optimization objectives. Increasing hardness may improve grinding efficiency but increases the risk of scratching the steel ball surface; improving wear resistance may require increasing alloying elements but increases material costs.

[0042] Constraints are set for the optimization objectives, including minimum requirements for the surface quality of the steel balls as hard constraints, and material cost and feasibility of the manufacturing process as soft constraints. These multiple objective functions and constraints are integrated into a standard multi-objective optimization problem formulation, resulting in a multi-objective optimization function for the grinding plate quality index.

[0043] Step 4.2: Obtain the candidate optimization solution set for the chemical composition of the grinding plate based on the genetic algorithm; Based on the multi-objective optimization function obtained in step 4.1 and the multi-modal quality index prediction model obtained in step 3, a genetic algorithm is used to perform a global search in the chemical composition parameter space to obtain a candidate optimal solution set for the chemical composition of the grinding plate. Specifically, the optimization variables are defined as the chemical composition parameters of the grinding plate, including the mass fractions of major elements such as carbon, silicon, manganese, phosphorus, and sulfur, as well as the mass fractions of optional alloying elements such as nickel, chromium, and molybdenum. Based on the composition range of gray cast iron and engineering experience, upper and lower bounds are set for the values ​​of each element to ensure that the generated formula is physically feasible and meets the casting process requirements. The population of the genetic algorithm is initialized by randomly generating a certain number of individuals, each representing a set of chemical composition formulas. The individual is encoded as a real number vector, and each component of the vector corresponds to the mass fraction of each element. The population size is set to 50 to 200 individuals, as the population size affects the search capability and computational cost of the algorithm. The fitness of each individual in the population is evaluated, and the individual's chemical composition vector is combined with hypothetical or typical microstructure characteristics and input into the multi-modal quality index prediction model in step 3 to obtain the quality index prediction results corresponding to the formula.

[0044] Based on the predicted results of quality indicators, the scores of individuals on each optimization objective are calculated. For multi-objective optimization, the Pareto dominance relation is used to evaluate the quality of individuals. If individual A is no worse than individual B on all objectives and is better than individual B on at least one objective, then A is said to dominate B. Individuals in the population are ranked and classified according to the Pareto dominance relation. Non-dominated individuals are assigned to the first frontier. After removing individuals from the first frontier, the remaining non-dominated individuals are assigned to the second frontier, and so on. For individuals within the same frontier, the crowding distance is calculated as a distinguishing metric. Individuals with large crowding distances have fewer surrounding individuals, and retaining individuals with large crowding distances helps maintain the diversity of the solution set. Based on fitness ranking and crowding distance, tournament selection or roulette wheel selection methods are used to select parent individuals for reproduction. Crossover operations are performed on the selected parent individuals, using simulated binary crossover or arithmetic crossover methods to generate offspring individuals. Crossover operations allow the superior traits of the parents to be combined and inherited. Mutation operations are performed on the offspring individuals, randomly perturbing certain genes of the individuals with a certain probability. Mutation operations introduce new genetic information and prevent the algorithm from prematurely converging to a local optimum. The magnitude and probability of mutation need to be carefully set. Excessive mutation destroys existing superior individuals, while insufficient mutation leads to stagnation in the search. Offspring individuals are added to the population, and the best-fitting individuals are selected from both parents and offspring according to an elite retention strategy to form a new generation, ensuring that superior individuals are not lost during evolution. The evolutionary cycle of selection, crossover, mutation, and elite retention is repeated, and after several generations, the population gradually converges towards the Pareto front. A termination condition is set: evolution stops when the preset maximum number of iterations is reached or the population's fitness no longer improves. Non-dominant individuals are extracted from the final population; these individuals constitute the Pareto optimal solution set. Each solution represents a set of optimal formulations that achieve different trade-offs among multiple objectives, resulting in a candidate optimal solution set for the grinding plate chemical composition.

[0045] Step 4.3: Obtain the optimization scheme of heat treatment process parameters based on Bayesian optimization method; Based on the candidate optimization solution set of the grinding plate chemical composition obtained in step 4.2, the Bayesian optimization method is used to optimize the heat treatment process parameters corresponding to each formulation, resulting in a complete set of optimization schemes. Specifically, for each chemical composition formulation in the candidate solution set, the heat treatment process parameters become variables that need further optimization, including annealing temperature, holding time, cooling rate, etc. The quality index prediction model of step 3.3 is extended by adding the heat treatment process parameters as additional input features to the model, retraining or fine-tuning the model so that it can predict the final performance of the grinding plate under given chemical composition and process parameters. Bayesian optimization is a suitable optimization method for objective function evaluation, which is costly. It guides the search by constructing a probabilistic surrogate model of the objective function. A Gaussian process regression model is constructed as a surrogate model of the objective function. The Gaussian process model can model the objective function based on a limited number of observation points and provide the predicted mean and predicted variance of the function value. In the initialization phase, several points are randomly sampled in the process parameter space. The objective function value is evaluated for each point using the quality index prediction model. The sampled points and function values ​​are used as training data for the Gaussian process model to fit the surrogate model. A data collection function is defined to balance the trade-off between exploration and utilization. Commonly used data collection functions include expected improvement, probabilistic improvement, or confidence upper bound. The data collection function calculates the value of each candidate point based on the predicted mean and uncertainty of the surrogate model; points with high predicted values ​​or high uncertainty are considered more valuable. The next process parameter point to be evaluated is selected by optimizing the data collection function. At this point, the true objective function is evaluated, i.e., the quality index is calculated using the quality index prediction model. The new observation point is added to the training data to update the Gaussian process surrogate model. This process is iterated, selecting a new point for evaluation and updating the surrogate model in each iteration. After several iterations, the surrogate model's approximation of the objective function becomes increasingly accurate, and the search process gradually converges to the optimal solution. Because Bayesian optimization fully utilizes existing evaluation results, it can find the optimal solution with fewer evaluations compared to random search or grid search, making it particularly suitable for scenarios where the computational cost of the quality index prediction model is high. For multi-objective optimization problems, a multi-objective Bayesian optimization method is adopted, where the data collection function considers the comprehensive performance of all objectives to find the Pareto optimal solution set. After optimization, for each group of chemical composition formulations, the corresponding optimal heat treatment process parameters are obtained. These parameters are combined to form a formulation-process synergistic optimization scheme, resulting in a complete set of optimization schemes that includes chemical composition and heat treatment parameters.

[0046] Step 4.4: Obtain the recommended optimal implementation scheme based on engineering constraints and cost analysis; Based on the complete set of optimized solutions obtained in step 4.3, engineering feasibility analysis and cost-benefit assessment are used to obtain the grinding plate preparation scheme. Specifically, the engineering feasibility of each scheme in the optimized scheme set is checked to assess whether the chemical components in the formula can be obtained through existing raw material procurement channels, whether the amount of alloying elements added is within the processing capacity of the smelting equipment, and whether the heat treatment process parameters are within the operating range of the existing heat treatment equipment. Schemes that do not meet the engineering constraints are removed from the candidate set to ensure that the final recommended scheme can be implemented in actual production. Cost estimation is performed on the retained schemes, calculating the raw material cost of each formula. The addition of expensive alloying elements will increase the material cost. The energy consumption cost of the heat treatment process is calculated, with high-temperature long-term heat preservation processes consuming even more energy. The service life of the grinding plate and the number of steel balls that can be processed during this service life are estimated, and the grinding plate cost per unit of steel ball is calculated. The material cost and preparation cost are allocated to the steel balls processed during the service life of the grinding plate. Taking into account both performance and cost, the cost-effectiveness index is calculated. Cost-effectiveness is defined as the ratio of performance benefits to cost. Performance benefits can be measured by the economic value brought by the improvement of steel ball quality or the time cost saved by the improvement of grinding efficiency. In the Pareto optimal solution set, the final implementation plan is selected based on the decision-maker's preferences. If the decision-maker prioritizes steel ball quality, the plan with the best performance in quality indicators is chosen; if cost control is a priority, the plan with the highest cost-effectiveness is selected; if a balance is required, a compromise between quality and cost is chosen. Multiple alternative plans are provided for decision-makers to choose from, each with detailed specifications, including chemical composition formulation, target mass fractions of each element and allowable deviation ranges, heat treatment process parameters including annealing temperature, holding time, cooling method, predicted quality indicators and their uncertainty range, material cost estimates, and expected service life. The recommended plan is output in the form of a technical document, including a formula table, process flow diagram, quality indicator prediction report, cost analysis table, etc., resulting in a grinding plate preparation plan that can be directly used by the production department.

[0047] The simulation evaluation module, based on the grinding plate preparation scheme, simulates the preparation process and records the simulation data and results to conduct a feasibility evaluation and obtain the scheme feasibility evaluation results. Step 5.1: Obtain a digital twin simulation model of the grinding process based on multiphysics modeling; Based on the physical process of bearing steel ball grinding, a digital twin model of the grinding process is constructed using a multiphysics coupling modeling method, resulting in a virtual simulation system capable of simulating the actual grinding process. Specifically, geometric models of the grinding plate and steel ball are established. The grinding plate is modeled as a disc-shaped structure with a certain curvature and surface texture, while the steel ball is modeled as a standard sphere. The relative motion relationship between the grinding plate and the steel ball is set according to the structural parameters of the actual grinding equipment. Material models are established to describe the mechanical properties of the grinding plate and steel ball. The material parameters of the grinding plate are set based on the chemical composition and predicted mechanical properties given by the optimization scheme, including elastic modulus, Poisson's ratio, hardness, and yield strength. The material parameters of the steel ball are set according to the standard performance data of bearing steel. A contact mechanics model is established to describe the contact behavior between the grinding plate and the steel ball. Hertz contact theory or finite element contact algorithms are used to calculate the contact stress distribution. The magnitude and distribution of contact stress directly affect material removal and surface quality formation. A wear model is established to describe the material removal mechanism during the grinding process. This model can employ Arcard's wear law or a more complex abrasive wear model. Model parameters are related to the material's hardness, toughness, and coefficient of friction. The wear model calculates the amount of material removed from the steel ball surface and the wear of the grinding plate per unit time. A friction model is established to describe the frictional behavior at the contact interface. The coefficient of friction is determined based on the material combination of the grinding plate and steel ball, as well as the lubricating effect of the grinding fluid. Friction affects energy dissipation and heat generation during the grinding process. A heat conduction model is established to describe the evolution of the temperature field during grinding. The heat generated by grinding friction is conducted and diffused through the grinding plate, steel ball, and grinding fluid. The distribution of the temperature field affects the mechanical properties of the material and its wear behavior. A surface morphology evolution model is established to describe the dynamic changes in the surface roughness and sphericity of the steel ball. The evolution of the surface morphology depends on the spatial uniformity of material removal and the cumulative effect over time. These physical models are coupled: the contact stress output from the contact mechanics model is used as the input to the wear model; the material removal output from the wear model is used as the basis for updating the geometric model; and frictional heat generation is used as the heat source input to the heat conduction model. The temperature field, in turn, affects the mechanical property parameters of the material. The coupled multiphysics model was numerically solved using the finite element method (FEM). The geometric regions of the grinding plate and steel ball were discretized into finite element meshes, and the governing equations were solved at the mesh nodes. A time-stepping method was used to simulate the dynamic evolution of the grinding process. Simulation boundary conditions and initial conditions were set. Boundary conditions included the grinding plate rotation speed, grinding pressure, and the flow state of the grinding fluid. Initial conditions included the initial surface morphology of the steel ball and the initial state of the grinding plate. Simulation parameters were calibrated by adjusting empirical parameters in the model using measurement data from actual grinding experiments, ensuring the simulation results matched the actual situation, resulting in a validated digital twin simulation model of the grinding process.

[0048] Step 5.2: Obtain the input configuration of the simulation model based on the optimized scheme parameters; Based on the grinding plate preparation scheme obtained from the formula optimization module, a parameter mapping method is used to convert the grinding plate formula and quality index prediction results in the optimized scheme into input parameters for the digital twin simulation model of the grinding process, thus obtaining the simulation model configuration corresponding to the optimized scheme. Specifically, the chemical composition and heat treatment process parameters of the grinding plate are extracted from the optimized scheme, and these parameters are input into the multimodal quality index prediction model in step 3 to obtain the quality index prediction results corresponding to the optimized scheme, including mechanical quality indices such as hardness, elastic modulus, and wear resistance. The predicted mechanical quality indices are mapped to material parameters in the simulation model. The hardness value is converted into yield strength and compressive strength using empirical formulas, and the wear resistance index is converted into the wear coefficient in the wear model. For the uncertainties contained in the prediction results, the variation range of the parameters is set in the simulation, and sensitivity analysis is performed to evaluate the impact of parameter fluctuations on the simulation results. The process parameters of the simulation are set, including grinding pressure, rotation speed, grinding time, etc., which are determined according to the processing technology standards of the target steel ball. Configure the simulation output monitoring items, specifying the physical quantities to be monitored, including the material removal rate on the steel ball surface, contact stress distribution, surface temperature, surface roughness evolution curve, sphericity error variation curve, and cumulative wear of the grinding plate. Set the simulation time span and time step, ensuring the time span covers a complete grinding cycle and the time step is small enough to capture detailed changes in the physical process. Check the completeness and rationality of the simulation configuration, ensuring all necessary input parameters are assigned values ​​and within physically reasonable ranges, thus obtaining a complete simulation model input configuration.

[0049] Step 5.3: Obtain the virtual grinding process results of the optimized scheme based on simulation calculations; Based on the simulation model input configuration obtained in step 5.2 and the digital twin simulation model of the grinding process constructed in step 5.1, a high-performance computing method is used to run the simulation calculation, obtaining detailed result data of the virtual grinding process corresponding to the optimized scheme. Specifically, the simulation calculation is started, and the solver calculates the distribution of physical quantities such as contact force, material removal, and temperature change at each time step according to the set control equations and boundary conditions. The simulation calculation starts from the initial state and gradually advances the simulation time, updating the surface morphology of the steel ball, the wear state of the grinding plate, and the temperature field of the system at each time step. The time history data of each monitoring item during the simulation are recorded, including the decrease curve of the steel ball surface roughness with grinding time, the convergence curve of the sphericity error with grinding time, and the increase curve of the cumulative wear of the grinding plate with grinding time. The distribution characteristics of contact stress are analyzed to identify whether there are stress concentration areas, which may lead to local damage or scratches on the surface of the steel ball. The distribution of the temperature field is analyzed to evaluate the highest temperature and temperature rise rate during the grinding process, as excessively high temperatures may cause changes in the material structure or thermal damage. At the end of the simulation, the final surface morphology data of the steel balls are extracted, and the surface roughness and sphericity accuracy of the steel balls after processing are calculated to assess whether they meet the quality requirements. The efficiency indicators of the grinding process are statistically analyzed, including the grinding time required to achieve the target surface quality, the total material removal amount, and energy consumption. The wear condition of the grinding plate is evaluated, the wear amount caused to the grinding plate by a single steel ball grinding cycle is calculated, and the service life of the grinding plate at this wear rate is estimated, i.e., the total number of steel balls that can be processed. All simulation output data are organized and visualized to generate contact stress cloud maps, temperature field distribution maps, surface morphology evolution animations, quality indicator time curves, etc., resulting in an intuitive virtual grinding process result report.

[0050] Step 5.4: Based on the simulation results analysis, obtain the feasibility assessment conclusion of the optimization scheme; Based on the virtual grinding process results obtained in step 5.3, a comprehensive analysis using engineering criteria and quality standards is conducted to assess the feasibility and expected effects of the optimized scheme in practical applications. Specifically, the simulated steel ball surface quality indicators are compared with the target quality standards to determine whether the optimized scheme can ensure that the processed steel balls meet the requirements for surface roughness and sphericity accuracy. If the simulation results show that all steel ball quality indicators meet the standards with a certain margin, the optimized scheme is highly feasible and can be recommended for implementation. If the simulation results show that some quality indicators are close to the critical value or exceed the standard, the reasons need to be analyzed and improvement suggestions proposed, which may require adjusting process parameters or re-optimizing the grinding plate formula. Potential risks discovered in the simulation are analyzed. If there is non-uniformity in the contact stress distribution, it may lead to ripples or local defects on the surface of the steel balls during actual processing, requiring measures to improve contact conditions or optimize the surface structure uniformity of the grinding plate. If the temperature field shows local high-temperature areas, the degree of thermal influence needs to be assessed, and the grinding speed may need to be adjusted or cooling measures enhanced if necessary. The service life of the grinding plate is evaluated. The number of steel balls that the grinding plate can process is estimated based on the wear rate calculated by simulation, and compared with economic requirements to determine whether the grinding plate service life meets production needs. If the grinding plate life is too short, the weight of wear resistance needs to be increased in the optimization objectives, and the optimization should be repeated. The grinding efficiency should be evaluated, and the production cycle time estimated based on the processing time obtained from simulation to determine whether it can meet the capacity requirements of the production plan. The simulation results of different optimization schemes should be compared, and the scheme with the best overall performance should be selected from multiple candidate schemes. A decision support report should be generated, containing detailed parameters of the optimization scheme, key data from the simulation results, feasibility analysis conclusions, risk warnings, and implementation suggestions, providing comprehensive information support for decision-makers and obtaining the scheme feasibility assessment results. When the scheme feasibility assessment results indicate that the grinding plate preparation scheme output by the formula optimization module does not meet the requirements, the assessment results should be fed back to the formula optimization module for re-optimization of multiple objectives until a grinding plate preparation scheme that meets the requirements is obtained. When the assessment conclusion confirms the feasibility of the scheme, it provides guidance for the actual preparation and application of the grinding plate.

[0051] The state monitoring module, based on the feasibility assessment results of the scheme, carries out grinding plate preparation according to the grinding plate preparation scheme, monitors the preparation process signals in real time and performs signal processing to obtain the comprehensive state feature vector of the grinding plate; Step 6.1: Acquire the raw monitoring signals of the grinding process based on the arrangement of multiple sensors; Based on the grinding plate preparation scheme and its feasibility assessment, the grinding plate was prepared according to the scheme. Based on the actual grinding equipment and process monitoring requirements, multiple types of sensors were used to comprehensively monitor the grinding process, obtaining raw monitoring signals containing multi-dimensional information including mechanics, vibration, temperature, and acoustics. Specifically, force sensors were installed at key locations on the grinding equipment to measure the normal pressure and tangential friction between the grinding plate and the steel ball during the grinding process. The amplitude and fluctuation characteristics of the force signals reflect the grinding capacity and surface condition of the grinding plate. Vibration sensors were installed on the grinding plate support structure to collect vibration signals during the grinding process. The vibration signals contain rich frequency components; different frequency components correspond to different physical processes. Low-frequency vibrations may be related to the imbalance of the grinding plate or the support stiffness, while high-frequency vibrations may be related to abrasive cutting and surface roughness. Temperature sensors were installed near the grinding plate and steel ball to monitor temperature changes in the grinding area. Temperature increases reflect the generation and dissipation of frictional heat; abnormal temperature increases may indicate poor lubrication or accelerated wear. Acoustic emission sensors are installed to capture acoustic emission signals generated by the microscopic damage of the material during the grinding process. These signals are sensitive to early damage such as crack initiation and abrasive grain shedding, providing early warning of changes in the grinding plate's condition. Displacement sensors monitor the thickness change of the grinding plate; the cumulative wear is measured by the reduction in thickness. The output signals of all sensors are synchronously acquired through a data acquisition system. An appropriate sampling frequency is set, which should be at least twice the highest frequency component of the signal to satisfy the Nyquist sampling theorem and avoid frequency aliasing. The acquired multi-channel time-series signals are correlated with timestamps, recording the acquisition time corresponding to each data point to ensure time alignment of data from different sensors. The raw monitoring signals are stored in a database, establishing associations between signals and grinding batches, grinding plate numbers, and process parameters, facilitating subsequent data retrieval and analysis, resulting in a multi-sensor raw monitoring signal database for the grinding process.

[0052] Step 6.2: Obtain the denoised purified signal based on adaptive Kalman filtering; Based on the original monitoring signal obtained in step 6.1, an adaptive Kalman filter algorithm is used to filter the signal in real time to obtain the denoised grinding process monitoring signal. Specifically, for each sensor channel, a state-space model of the signal is established. The state equation describes the time evolution of the true signal value, and the observation equation describes the relationship between the sensor measurement value and the true value. The measured value equals the true value plus the measurement noise. The Kalman filter estimates the true signal value from the noise-contaminated measurement value through a recursive prediction and update process. The prediction step predicts the current state based on the state estimate of the previous time step and the system model. The update step uses the measurement value at the current time step to correct the prediction result. The correction weight is determined by the Kalman gain, which is automatically calculated based on the prediction error covariance and the measurement noise covariance. Traditional Kalman filtering assumes that the statistical characteristics of system noise and measurement noise are known and constant. However, in actual industrial environments, noise characteristics change with operating conditions, and the performance of a fixed-parameter filter will degrade. The adaptive Kalman filter method is used to estimate the statistical characteristics of noise in real time during the filtering process and dynamically adjust the filter parameters based on the estimation results. The noise covariance is estimated based on the statistical characteristics of the innovation sequence, i.e., the prediction error sequence. The variance of the innovation sequence reflects the combined level of measurement noise and model uncertainty. The sample variance of the innovation sequence is calculated using a sliding window, and the noise covariance matrix is ​​updated using the sample variance. When a sudden change in the statistical characteristics of the innovation sequence is detected, it indicates a change or anomaly in the system operating conditions. The adaptive mechanism quickly adjusts the filtering parameters to adapt to the new state. Adaptive Kalman filtering is performed on the signal of each sensor channel to obtain the denoised signal estimate and the estimated error covariance. The estimated error covariance characterizes the reliability of the filtering result. The quality of the filtered signal is evaluated, the signal-to-noise ratio improvement factor is calculated, and the filtering effect is verified. The filtered purified signal is used as the input for subsequent analysis to obtain the denoised grinding process monitoring signal.

[0053] Step 6.3: Obtain robust comprehensive state features based on multi-sensor data fusion; Based on the denoised grinding process monitoring signal obtained in step 6.2, multi-sensor data fusion technology is used to integrate information from different physical quantities to obtain more robust and comprehensive grinding plate state characteristics. Specifically, feature extraction is performed on the signals from each sensor. Statistical features are extracted from the time-domain signal, including the mean reflecting the steady-state level of the signal, the standard deviation reflecting the degree of signal fluctuation, the peak value and peak-to-peak value reflecting the extreme cases of the signal, and the skewness and kurtosis reflecting the shape characteristics of the signal distribution. Frequency domain analysis is performed on the signal. The time-domain signal is converted into a frequency domain representation through Fast Fourier Transform, and spectral features are extracted, including the dominant frequency, spectral energy distribution, and the energy proportion of specific frequency bands. Spectral features can reveal the periodic components and frequency structure changes in the signal. Time-frequency analysis is performed on the signal. Wavelet transform or short-time Fourier transform is used to obtain the time-frequency graph of the signal. Time-frequency features can capture the frequency variation of the signal over time and are suitable for analyzing non-stationary signals. The feature vectors extracted from different sensors are concatenated to form a high-dimensional multimodal feature vector. Feature-level fusion is performed using dimensionality reduction methods such as principal component analysis or linear discriminant analysis to project high-dimensional feature vectors onto a low-dimensional feature space, removing redundancy between features and retaining the most discriminative principal components. Decision-level fusion is then performed. For classifying the grinding plate state, different sensors independently provide state judgment results based on their respective signals. Voting, weighted averaging, or Bayesian fusion methods are used to synthesize multiple judgment results to obtain the final decision. Fusion decision is more reliable than single-sensor judgment. A consistency verification mechanism is established between sensors. By analyzing the correlation between signals from different sensors, abnormal sensors are identified. If the signal of a certain sensor does not show a consistent trend with other sensors, the sensor may be faulty or subject to local interference. Data from abnormal sensors are removed or their weight in the fusion is reduced. For missing or abnormal sensor data, estimation and compensation are performed using data from other sensors and historical correlations to ensure the continuity and integrity of state features. The fused comprehensive state features are used as input to the grinding plate state assessment and life prediction model to obtain the comprehensive grinding plate state feature vector.

[0054] The life prediction module constructs a quality index degradation prediction model, combines the comprehensive state feature vector of the grinding plate to perform real-time state assessment of the grinding plate, predicts the remaining service life of the grinding plate, and obtains the remaining service life prediction results and maintenance decision suggestions. Step 7.1: Train and obtain an initial quality indicator degradation prediction model based on historical data; Based on historically collected full lifecycle monitoring data and quality indicator degradation records of grinding plates, a time-series prediction model is constructed using a Long Short-Term Memory (LSTM) network to obtain a quality indicator degradation prediction model. Specifically, complete monitoring data of multiple grinding plates from their initial use to their replacement are compiled from the historical database. Each grinding plate corresponds to a time-series data sequence, which includes the comprehensive state feature vector at each moment, cumulative working time, cumulative number of processed steel balls, and measured values ​​of quality indicators. Quality indicators include the current hardness, surface roughness, grinding efficiency, and quality of the processed steel balls, etc., which gradually degrade with the use of the grinding plate. The time-series data is preprocessed and standardized, with data points arranged in chronological order, and features are normalized to ensure that different features are within the same numerical range. A LSTM network model is constructed, which is specifically designed to process time-series data and can learn long-term dependencies. The network consists of an input layer, multiple LSTM layers, and an output layer. The input layer receives state feature vectors from the current time and several historical time points. The LSTM layers selectively memorize and forget information through a gating mechanism, enabling them to capture long-term trends and short-term fluctuations in time-series data. The output layer predicts quality index values ​​at several future time points or directly predicts remaining lifespan. Training samples are designed by generating multiple training samples using a sliding window method for each complete time-series dataset. Each sample contains a historical observation window and a corresponding future prediction target. The length of the historical window is set according to the time scale of the grinding plate quality index degradation. A loss function is defined to measure the difference between the predicted and true values. For regression prediction tasks, mean squared error or mean absolute error is used; for lifespan prediction tasks, a dedicated survival analysis loss function can be used. The LSTM network is trained using a historical dataset, and the network parameters are updated using the time backpropagation algorithm, with appropriate learning rates and batch sizes set. To prevent overfitting, regularization techniques and an early stopping strategy are used; training stops when the loss on the validation set no longer decreases. After training, the model's predictive performance is evaluated on the test set, and the prediction error and accuracy are calculated to obtain the quality index degradation prediction model.

[0055] Step 7.2: Obtain the real-time evaluation result of the current state of the grinding plate based on the real-time data stream; Based on the quality index degradation prediction model obtained in step 7.1 and the comprehensive state feature vector of the grinding plate obtained by the state monitoring module, a model inference method is used to perform real-time state evaluation of the grinding plate currently in use, obtaining the real-time evaluation result of the grinding plate's current state. Specifically, the historical state feature sequence of the grinding plate from the start of use to the current moment is used as input. The input sequence includes information such as state features within a past time window, cumulative working time, and the number of steel balls processed. The input sequence is fed into a trained LSTM model for forward propagation calculation. Based on the evolution law of historical state features, the model infers the quality index of the grinding plate at the current moment, including the current actual hardness, wear resistance, grinding efficiency, and other performance parameters, as well as the degree of degradation of these performances relative to the initial state. The quality index degradation rate is calculated, defined as the difference between the current performance and the initial performance divided by the initial performance. The magnitude of the degradation rate reflects the degree of aging of the grinding plate. The grinding plate's condition is categorized into different levels based on the degree of quality indicator degradation. For example, a healthy state indicates that the quality indicators are within the normal range and require no attention; a watchful state indicates that performance is beginning to degrade significantly and requires close monitoring; a warning state indicates that the quality indicators are approaching the threshold and replacement is needed; and a fault state indicates severe performance degradation and immediate replacement is required. Based on the current performance evaluation results and preset state thresholds, the grinding plate's current state level is determined, and a state evaluation report is generated. The report includes the current values ​​of each quality indicator, the degree of degradation, the state level, and trend descriptions. The state evaluation results are displayed in real-time on the monitoring interface using dashboards, graphs, status lights, and other visual methods, allowing operators to intuitively understand the grinding plate's working status and obtain real-time evaluation results of its current condition.

[0056] Step 7.3: Obtain the prediction results of the remaining service life of the grinding plate based on the trend extrapolation method; Based on the real-time evaluation results of the current state of the grinding plate obtained in step 7.2 and the quality index degradation prediction model in step 7.1, a trend extrapolation method is used to predict the future quality index degradation trajectory of the grinding plate, obtaining the remaining service life prediction results and maintenance decision suggestions. Specifically, an LSTM model is used for multi-step prediction. Starting from the current moment, the quality index values ​​at each future moment are predicted iteratively, with the output of each prediction serving as the input for the next prediction, gradually advancing towards the future. A prediction curve of the quality index over time is generated, showing the decay trend of each quality index in the future. A performance failure threshold is set, which is the critical value for judging that the grinding plate no longer meets the usage requirements, such as a decrease in hardness exceeding a certain percentage, grinding efficiency falling below the minimum requirement, or the quality of the processed steel balls being unqualified. The moment when the quality index first reaches the failure threshold is found on the prediction curve, and the time difference between this moment and the current moment is the predicted remaining service life. Considering the uncertainty of the prediction, since the prediction model itself has errors, future operating conditions may differ from historical conditions, and the remaining service life prediction is uncertain. The probability distribution of the remaining service life is obtained by using the prediction variance output by the model or by performing multiple predictions using the Monte Carlo method, giving the expected value and confidence interval of the remaining service life. Based on the remaining service life prediction results, maintenance decision recommendations are generated. If the predicted remaining service life is sufficient, it is recommended to continue normal operation and maintain monitoring. If the remaining service life is close to the warning threshold, it is recommended to prepare spare parts in advance and arrange a replacement plan. If the remaining service life is very short, it is recommended to arrange a shutdown for replacement as soon as possible to avoid affecting production. Combining production plans and inventory status, the timing of replacement is optimized, selecting an appropriate time window for replacement while ensuring production continuity, avoiding emergency shutdowns and unplanned production stoppages. A maintenance decision report is generated, including the predicted remaining service life value and confidence interval, the recommended replacement time range, the required spare parts list, the expected replacement operation time, etc., yielding the remaining service life prediction results and maintenance decision recommendations.

[0057] Step 7.4: Obtain a continuously optimized prediction model based on an online learning mechanism; Based on the predicted remaining service life of the grinding wheel and the actual performance data after its use, an online learning method is used to incrementally update the prediction model, resulting in an adaptive model with continuously improving prediction accuracy. Specifically, during the use of the grinding wheel, offline performance tests are conducted periodically or at critical moments to obtain actual quality indicator measurements. Measurement timing includes regular maintenance checks, special inspections when abnormal signs are detected, and final evaluations before grinding wheel replacement. The actual measured values ​​are compared with the model's predicted values ​​at the same time, the prediction error is calculated, and the magnitude and direction of the error are analyzed to determine whether the model prediction is overly optimistic or conservative. Newly obtained measured data are used as new training samples, and the online learning algorithm is used to incrementally update the model. Online learning does not require retraining the entire model from scratch; instead, it maintains the learned knowledge and fine-tunes the model parameters based on new data, making the update process fast and efficient. An incremental gradient descent method is used, employing new samples to calculate the gradient of the loss function and make small adjustments to the model weights, with the adjustment step size controlled by the learning rate. To avoid overfitting or catastrophic forgetting caused by new data, regularization constraints based on historical knowledge are maintained during updates. New model parameters should not deviate too far from the original model unless the new data provides a strong update signal. A sliding window strategy is employed, retaining data from the most recent period for model updates, while outdated data is phased out over time, enabling the model to adapt to long-term operating conditions. Model performance on new data is monitored; if online updates improve performance, the update is confirmed and the new model is saved; if performance deteriorates, the new data may be anomaly-prone or insufficiently representative, requiring manual review before adoption. Through continuous online learning, the model gradually adapts to the characteristics of the current production environment, becoming increasingly accurate in predicting specific grinding batches and process conditions. A model version management mechanism is established, recording the update time, basis, and performance changes for easy traceability and rollback. The online learning mechanism is combined with the closed-loop feedback of the entire system. Actual data is not only used to optimize the prediction model but also fed back to the quality index prediction model in the prediction module and the optimization algorithm in the formulation optimization module, forming a continuous improvement cycle for the entire system, resulting in an adaptive quality index degradation prediction model with continuously improving prediction accuracy. The actual usage data obtained from online learning is fed back to the quality indicator prediction model of the prediction module and the optimization algorithm of the formula optimization module, forming a closed-loop system for continuous improvement.

[0058] In one embodiment of the present invention, a specific example is provided: A precision bearing manufacturer adopted the deep learning-based gray cast iron grinding plate microstructure and performance optimization system of this invention to comprehensively optimize its steel ball grinding process, achieving certain application results.

[0059] The company produces bearing steel balls with a diameter of 12.7 mm, made of GCr15 bearing steel, used in high-end CNC machine tool spindle bearings. The requirements for surface roughness, sphericity accuracy, and batch consistency of the steel balls are extremely high. The company's original gray cast iron grinding plates used a traditional, experience-based formula, with main chemical components of 3.2% carbon, 2.1% silicon, 0.7% manganese, 0.08% phosphorus, and 0.05% sulfur, processed using conventional annealing techniques. Due to fluctuations in the composition of the purchased cast iron raw materials between batches, even when prepared using the same formula and process, the performance of different batches of grinding plates varied, leading to unstable steel ball processing quality.

[0060] After introducing the system of this invention, the company conducted comprehensive data collection on its existing grinding plates. Metallographic images of 30 batches of grinding plates were collected, with 25 microscopic images taken from different fields of view for each batch, resulting in a total of 750 high-resolution metallographic images. Chemical composition and physical property tests were performed on each batch of grinding plates, establishing a complete database of grinding plate characteristics. Simultaneously, data on the grinding effect of these grinding plates in actual use were recorded, including the surface quality of the processed steel balls, grinding efficiency, and grinding plate lifespan.

[0061] The system performed in-depth analysis and modeling of the collected data. An improved convolutional neural network was used to automatically analyze metallographic images, identifying and quantifying the morphological characteristics of graphite and the matrix structure of the grinding plate. Analysis revealed that the graphite in the existing grinding plate was predominantly flake-like with uneven length distribution, and some areas contained coarse graphite flakes. The ferrite content in the matrix structure fluctuated significantly. A multimodal fusion Bayesian deep neural network model was constructed, establishing a full-chain prediction model from chemical composition and microstructure to macroscopic performance and grinding effect.

[0062] Based on the trained prediction model, the system performs multi-objective optimization of the grinding plate formulation and process parameters. The optimization objectives include maximizing grinding plate lifespan, maximizing grinding efficiency, minimizing adverse effects on steel ball surface quality, and controlling material costs. Through global search using a genetic algorithm and local fine-tuning using Bayesian optimization, the system provides optimized grinding plate formulation suggestions.

[0063] Table 1 shows examples of chemical composition formulation data for some optimized schemes: Table 1: Examples of chemical composition formulation data for some optimized schemes;

[0064] Note: The content values ​​in the table are in percentage by mass.

[0065] The company selected Solution A01 for small-batch trial production verification. Five batches of test grinding plates were prepared according to the optimized formula, and each batch underwent comprehensive performance testing and actual grinding experiments. Test results showed that the optimized grinding plates exhibited significantly improved hardness uniformity, finer and more uniform graphite morphology in the microstructure, and the ferrite content in the matrix was controlled within the target range.

[0066] Table 2 shows an example of the actual grinding effect data of the test grinding plate: Table 2: Examples of actual grinding effect data for the test grinding plates;

[0067] Note: The average grinding time for the original empirical formula grinding plate is 22 minutes, and the average number of steel balls processed is 6500.

[0068] After successful trial verification, the company will fully implement the optimized solution. The system also deploys real-time monitoring and prediction modules to monitor the condition and predict the lifespan of grinding plates in use. Force sensors, vibration sensors, and temperature sensors are installed on the grinding equipment to collect real-time data on the grinding plate's operating status. Based on a performance degradation prediction model using a long short-term memory network, the system can predict when a grinding plate needs replacement 3 to 5 days in advance, making maintenance plans more reasonable and avoiding quality fluctuations and production interruptions caused by sudden grinding plate failure.

[0069] After applying this invention's system for six months, the company experienced improved stability in steel ball processing quality, significantly reduced batch-to-batch quality fluctuations, decreased customer complaint rates, and maintained a high product qualification rate. The average lifespan of grinding plates was extended, reducing the cost of grinding plates per unit product. Increased grinding efficiency boosted production line capacity, meeting the demands of growing orders. The system's intelligent decision support reduced reliance on experience among process engineers, allowing new employees to quickly grasp the key points of grinding plate selection and usage.

[0070] The company also leveraged the system's continuous learning capabilities to constantly accumulate optimization experience. With the accumulation of more data, the system became increasingly adaptable to different raw material batches, quickly providing adjustment suggestions based on the actual composition of each batch, achieving true personalized optimization. The company extended the application experience of this system to the grinding process optimization of other specifications of steel balls, achieving similar positive results and validating the system's versatility and scalability.

[0071] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A deep learning-based system for optimizing the microstructure and properties of gray cast iron grinding plates for bearing steel balls, characterized in that, include: The data acquisition module collects multimodal feature data of the grinding plate sample, performs preprocessing and feature processing, and obtains a multimodal fusion sample dataset. The tissue feature extraction module, based on a multimodal fusion sample dataset, deeply extracts tissue features to obtain a micro-tissue feature vector dataset; The prediction module, based on the micro-organism feature vector dataset and the multimodal fusion sample dataset, constructs a multimodal quality indicator prediction model to predict quality indicators and obtain the quality indicator prediction results. The formulation optimization module, based on the quality index prediction results, performs synergistic optimization of the grinding plate material formulation and preparation process to obtain the grinding plate preparation scheme; The simulation evaluation module, based on the grinding plate preparation scheme, simulates the preparation process and records the simulation data and results to conduct a feasibility evaluation and obtain the scheme feasibility evaluation results. The state monitoring module, based on the feasibility assessment results of the scheme, carries out grinding plate preparation according to the grinding plate preparation scheme, monitors the preparation process signals in real time and performs signal processing to obtain the comprehensive state feature vector of the grinding plate; The life prediction module constructs a quality index degradation prediction model, combines the comprehensive state feature vector of the grinding plate to perform real-time state assessment of the grinding plate, predicts the remaining service life of the grinding plate, and obtains the remaining service life prediction results and maintenance decision suggestions.

2. The deep learning-based microstructure and property optimization system for gray cast iron grinding plates for bearing steel balls according to claim 1, characterized in that, The data acquisition module includes acquiring microstructure image data of the grinding plate sample, processing the grinding plate sample using a standard metallographic sample preparation process, cutting a sample block from a preset depth region below the working surface of the grinding plate, performing coarse grinding, fine grinding and polishing treatments in sequence, performing etching treatment with a nitric acid alcohol solution of preset concentration, and observing and photographing the sample under a metallographic microscope at multiple magnifications to obtain a raw database of metallographic structure images.

3. The deep learning-based microstructure and property optimization system for gray cast iron grinding plates for bearing steel balls according to claim 1, characterized in that, The data acquisition module also includes: Chemical composition data of the grinding plate sample was obtained, and the chemical composition was quantitatively detected by a spectrometer to obtain the mass fraction data of elements in the grinding plate. Sample blocks suitable for spectrometer analysis were cut from the grinding plate, and the sample surface was polished to remove the oxide layer and contaminants. The sample was excited using a direct-reading spectrometer, and the intensity of characteristic spectral lines was measured using the direct-reading spectrometer. The intensity of the characteristic spectral lines was converted into elemental content data. For each sample, the measurement was repeated at different positions for no less than a preset number of times, and the average value of the measurement results was taken as the chemical composition data of the sample.

4. The deep learning-based microstructure and property optimization system for gray cast iron grinding plates for bearing steel balls according to claim 1, characterized in that, The tissue feature extraction module includes: A metallographic feature extraction model was constructed by loading the network structure and weight parameters of a pre-trained convolutional neural network model, retaining the convolutional layer part of the pre-trained convolutional neural network model as a feature extractor, and removing the fully connected classification layer of the original model. A new layer structure is added to the top of the pre-trained convolutional neural network model, and a hierarchical freezing strategy is used for training. In the early stage of training, the parameters of the bottom convolutional layers of the pre-trained model are frozen, and only the newly added top layer structure is trained. In the later stage of training, some high-level convolutional layers are unfrozen, and some high-level convolutional layers are fine-tuned.

5. The deep learning-based microstructure and property optimization system for gray cast iron grinding plates for bearing steel balls according to claim 4, characterized in that, The tissue feature extraction module also includes: Design a deep convolutional neural network model for extracting tissue features. The model adopts a two-branch convolutional neural network structure. The first branch is used for the identification and quantification of graphite features, and the second branch is used for the identification and quantification of matrix tissue features. Within each branch, an attention mechanism module is introduced, which includes two types: spatial attention and channel attention. The spatial attention mechanism assigns different weights to different positions based on the spatial importance of the feature map, while the channel attention mechanism assigns different weights to different channels based on the importance of the feature channels. The feature vectors output by the two branches are then concatenated or weighted and fused.

6. The deep learning-based microstructure and property optimization system for gray cast iron grinding plates for bearing steel balls according to claim 1, characterized in that, The prediction module includes: Construct an initial quality indicator prediction model and design a multi-input multi-output fully connected deep neural network architecture, which includes two independent input branches; The first input branch receives the microscopic tissue feature vector, which consists of multiple fully connected layers. Each layer contains several neuron nodes, and the nodes are fully connected. Non-linearity is introduced by an activation function. The second input branch receives the chemical composition feature vector and consists of multiple fully connected layers. After each of the two input branches completes several layers of transformation, feature fusion is performed in the middle layer of the fully connected deep neural network.

7. The deep learning-based microstructure and property optimization system for gray cast iron grinding plates for bearing steel balls according to claim 6, characterized in that, The prediction module also includes: The Bayesianization of the initial quality index prediction model transforms the weight parameters in the initial quality index prediction model from fixed values ​​into random variables that follow a probability distribution. Each weight is described by a probability distribution with mean and variance, assuming that the prior distribution of the weights is a standard normal distribution. An approximate inference is performed using variational inference. It is assumed that the posterior distribution follows a diagonal Gaussian distribution. Each weight is described by two parameters: mean and variance. The reparameterization technique is used to make the gradient backpropagate through random sampling. The loss function of variational inference is defined.

8. The deep learning-based microstructure and property optimization system for gray cast iron grinding plates for bearing steel balls according to claim 1, characterized in that, The formula optimization module includes: The candidate optimization solution set for the chemical composition of the grinding plate is obtained based on the genetic algorithm. The optimization variable is defined as the chemical composition parameter of the grinding plate. The upper and lower bounds of the value of each element are set according to the composition range of the grinding plate material. The genetic algorithm population is initialized by randomly generating a preset number of individuals. Each individual represents a set of chemical composition formulas and is encoded as a real number vector. Each component of the vector corresponds to the quality score of each element. The fitness of each individual in the population is evaluated, and the Pareto dominance relation is used to evaluate the quality of individuals. An evolutionary cycle of selection, crossover, mutation, and elite retention is performed based on fitness ranking and crowding distance.

9. The deep learning-based microstructure and property optimization system for gray cast iron grinding plates for bearing steel balls according to claim 1, characterized in that, The simulation evaluation module includes: A digital twin simulation model of the grinding process is obtained based on multiphysics modeling; Establish geometric models for the grinding plate and the steel ball. The grinding plate is modeled as a disk-shaped structure with preset curvature and surface texture, and the steel ball is modeled as a standard sphere. Establish material models to describe the mechanical properties of the grinding plate and the steel ball. A contact mechanics model is established to describe the contact behavior between the grinding plate and the steel ball, and the contact stress distribution is calculated; a wear model is established to describe the material removal mechanism during the grinding process; a friction model is established to describe the friction behavior of the contact interface; a heat conduction model is established to describe the temperature field evolution during the grinding process; and a surface morphology evolution model is established to describe the dynamic changes in the surface roughness and sphericity of the steel ball. The geometric model, material model, contact mechanics model, wear model, friction model, heat conduction model, and surface morphology evolution model are coupled.

10. The deep learning-based microstructure and property optimization system for gray cast iron grinding plates for bearing steel balls according to claim 1, characterized in that, The lifetime prediction module includes: A quality indicator degradation prediction model was built and trained based on historical data; Based on the quality index degradation prediction model and the comprehensive state feature vector of the grinding plate, the current state of the grinding plate is evaluated in real time, and the real-time evaluation result of the current state of the grinding plate is obtained. Based on the real-time evaluation results of the current state of the grinding plate and the quality index degradation prediction model, the trend extrapolation method is used to predict the future quality index degradation trajectory of the grinding plate, and obtain the remaining service life prediction results and maintenance decision suggestions.

Images