A method, equipment and medium for predicting the roughness of cylinder bore honing.

Abstract

This invention discloses a method, device, and medium for predicting the surface roughness of cylinder bore honing, belonging to the field of precision machining quality prediction technology. It addresses the problem of improving the prediction accuracy of cylinder bore honing surface roughness. The invention introduces a deep cross-network and utilizes vector outer product operations to construct high-order combined features layer by layer, explicitly capturing the nonlinear physical coupling effect between process parameters such as honing pressure and reciprocating speed. Furthermore, it enhances the diversity of statistical noise through a random partitioning strategy for virtual subbatches, reducing memory overhead while improving model generalization ability. This effectively overcomes the problem of weak nonlinear fitting ability in traditional methods, significantly improving the accuracy of model prediction under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and precision machining quality prediction technology, and relates to a method, equipment and medium for predicting the roughness of cylinder bore honing based on multi-level feature interaction and continuous attention. Background Technology

[0002] With increasing demands on the friction, sealing, and durability of internal combustion engines, the microstructure of the cylinder bore surface has become a key indicator affecting its overall performance. Honing, as a core finishing process, directly determines the final surface roughness through its process parameters. Traditional parameter setting methods relying on trial and error or empirical formulas are costly and time-consuming, making them unsuitable for flexible production needs involving small batches and diverse product varieties. Therefore, establishing a precise mapping model between process parameters and surface roughness to achieve advanced prediction of processing quality is of great significance for shortening the trial production cycle and reducing the scrap rate.

[0003] In recent years, data-driven surface roughness prediction methods have gradually become a research hotspot and have been applied to machining quality prediction. Traditional machine learning methods such as SVR, RF, and ANN can establish a basic mapping relationship between process parameters and roughness, but they are essentially shallow models. When faced with the conditions of strong multivariate coupling and small sample labeled data in actual honing processes, they generally suffer from insufficient nonlinear fitting ability and poor generalization performance. Furthermore, these methods typically rely on manual extraction of time-frequency domain statistical features, making it difficult to fully utilize all the information in the original data and easily leading to the loss of key features.

[0004] Based on this, deep learning technology has gradually been introduced into the field of machining quality prediction. Existing technology, such as the invention patent with publication number CN121167220A, discloses a deep learning-based method for predicting surface roughness in precision boring and milling machines. This method combines a gated recurrent unit (GRU) with an attention mechanism to construct a prediction model. The GRU extracts temporal features of machining parameters, and the attention mechanism assigns weights to different input features at different times, thus improving prediction accuracy to some extent. However, existing technologies still have the following shortcomings: To improve prediction accuracy, existing deep learning methods often require the construction of complex black-box models with a large number of parameters, which are prone to overfitting under small sample conditions. At the same time, due to the lack of effective feature focusing mechanisms and regularization constraints, the models are difficult to output decision-making basis that can be understood by process engineers. Although deep learning technology has strong expressive power, it is inefficient in processing structured tabular data and lacks key feature focus mechanisms, which limits the interpretability of the models.

[0005] Although some existing technologies have improved model transparency to some extent by introducing sparse attention mechanisms into the TabNet model, the hard threshold function they use is prone to excessive sparsity, leading to the omission of key features. In the honing process, there are complex physical coupling effects between various process parameters. Existing technologies mainly rely on the implicit feature combination of stacked fully connected layers or recurrent neural networks to approximate these interactions, failing to fully explore the high-order nonlinear coupling relationships between input process variables and lacking explicit feature cross-modeling mechanisms. This makes it difficult for the model to fully capture the deep coupling effects between parameters, thus limiting the prediction accuracy.

[0006] Furthermore, existing methods often neglect measurement noise, abnormal fluctuations, and spatial inconsistencies during data preprocessing, severely impacting modeling stability. Moreover, the networks generally lack effective feature recalibration mechanisms to enhance important feature representations. Therefore, there is an urgent need in this field for a high-precision, robust, and interpretable method for predicting the surface roughness of cylinder bore honing surfaces to overcome existing technological bottlenecks and meet the demands of intelligent manufacturing for accurate prediction of processing quality and process optimization. Summary of the Invention

[0007] The technical problem to be solved by this invention is how to improve the prediction accuracy of surface roughness prediction for cylinder bore honing.

[0008] The present invention solves the above-mentioned technical problems through the following technical solutions: A method for predicting the roughness of cylinder bore honing includes the following steps: Determine the input features, obtain surface roughness data of cylinder bore honing under different process combinations, and determine the target output vector; The input features are batch normalized in the initialization layer, and the cross product of the normalized features and the current layer state vector is calculated to construct higher-order combined features layer by layer. Multi-layer nonlinear mapping is performed on the higher-order combined features, and the first output feature after nonlinear mapping is divided into an initial backbone feature vector and an initial attention guidance signal along the feature dimension. The initial attention guidance signal is input into the encoding layer, a feature selection mask is generated in each decision step, and a new backbone feature vector and attention guidance signal are generated by combining the input features; the new attention guidance signal is used for the next decision step, and the new backbone feature vector is input into the fully connected prediction layer. The fully connected prediction layer performs nonlinear activation and aggregation on the backbone feature vectors output from multiple decision steps, maps them to the target space, and generates the predicted target output vector.

[0009] Furthermore, the input features include the honing reciprocating speed. , fine honing speed Jingheng expansion force Honing feed rate Platform honing reciprocating speed Platform honing speed Honing pressure and number of processing times ; The target output vector includes the maximum contour height. Peak height Core roughness depth Valley depth First material ratio Second material ratio ; Orthogonal honing experiments were conducted, and measurement points were arranged at the upper, middle and lower sections along the cylinder bore axis to perform multiple roughness measurements and collect the original surface roughness data. For the measurement results of the same index in each group of experiments, the threshold range for judging outliers for each index is determined by using the interquartile range as the interval. If the value exceeds the range, it is considered an abnormal measurement value and is removed. Calculate the coefficient of variation (CV) of the remaining valid measurements to assess spatial stability, and calculate the cross-sectional difference index between the upper, middle, and lower sections, including the absolute difference between the sections. Relative proportions of cross sections ; The arithmetic mean of the effective measured values ​​that meet the preset requirements of coefficient of variation and cross-sectional difference index in each group of experiments is taken as the final roughness characterization value under the process combination. After the above preprocessing, multiple sample pairs were formed in each group of experiments. ,in For the first The input feature vector of the group experiment, For the first The preprocessed target output vector of the group experiment is used to divide all samples into training dataset and test dataset according to the proportion.

[0010] Furthermore, the process of constructing the higher-order combined features is as follows: In the The layer uses the following logic to perform cross-calculation to construct higher-order combined features. :

[0011] in, The input features are after batch normalization. This represents the vector outer product operation. For the first The current state vector of the layer, and These are the learnable weight matrix and bias term for the corresponding layer, respectively.

[0012] Furthermore, the process of generating the initial backbone feature vector and the initial attention guidance signal is as follows: First, the higher-order combined features are input into the first feature transformer for nonlinear transformation to obtain the output features. ; wherein, the output feature Using the following logical representation:

[0013] in, It is a high-order combinatorial feature. b is the batch size. The dimension of the input feature. It is the set of real numbers; Choose a mask for the first feature, and take... ; For element-wise multiplication; The first feature processing function consists of several fully connected layers, a batch normalization module, a GLU activation function, and residual connections. The calculation process is as follows:

[0014] in, It is a high-order combination feature vector The result after linear transformation , For learnable weight matrix, For bias terms; It is the Sigmoid activation function. Output for linear path ; Next, output features The initial backbone features are separated using the Split module. With initial attention guidance signals This can be represented using the following logic:

[0015] in, This is the initial backbone feature vector. As an initial attention-guiding signal, This is the operation function for segmenting along the feature dimension.

[0016] Furthermore, the process of generating the new backbone feature vector and attention guidance signal is as follows: In each decision-making step, the attention transformer receives the attention guidance signal output from the previous step, sequentially passes it through a fully connected layer for linear mapping and batch normalization, and then combines it with a learnable historical attention intensity adjustment vector. Generate feature selection masks in continuous real number form The attention transformer uses a Softmax attention mechanism to generate a feature selection mask for the th... Step Feature Selection Mask This can be represented using the following logic:

[0017] in, It is a standardized exponential function; It is a mapping function consisting of a fully connected layer and a batch normalization module; This serves as an attention guidance signal for the previous step. This indicates element-wise multiplication; The historical attention intensity adjustment vector is learned from the previous step. The historical attention intensity adjustment vector is dynamically updated with the number of decision steps, and the attention priority of the corresponding channel is gradually reduced based on the mask selected by the previously generated features. The masking module performs element-wise multiplication of the generated feature selection mask with the batch-normalized input features, and outputs the masked features. ; Features after masking Input the feature transformer of the current step to perform nonlinear transformation and output features Then, after the Split operation, the main features of the current step are output. With new attention-guiding signals ; wherein, the output feature Using the following logical representation:

[0018] in, The input features are after batch normalization. For the first Feature selection mask for each decision step; This indicates element-wise multiplication; For the first The feature processing function in each decision step consists of several fully connected layers, a batch normalization module, a GLU activation function, and residual connections. Based on the set total number of steps After repeating the above process cyclically, all the main features will be output. polymerization.

[0019] Furthermore, the historical attention intensity adjustment vector It updates dynamically with each decision step, and its update rules are as follows:

[0020] in: For the first Feature selection mask generated in each decision step ; The relaxation factor is an adjustable hyperparameter; when any feature has been frequently selected in the early stages, the corresponding channel's... The value gradually decreases.

[0021] Furthermore, the process of generating the predicted surface roughness value for the cylinder bore honing is as follows: An accumulation operation with a non-linear activation function is used to aggregate the backbone features output from each decision step to obtain the global decision features output from the encoding layer. The polymerization process is represented as:

[0022] in, Let ∑ be the linear rectified activation function, and let ∑ be the accumulation operation. For the first The main features output by each decision step This represents the total number of decision-making steps. The aggregated global decision features are linearly projected into the target parameter space to obtain the surface roughness prediction vector for cylinder bore honing. The mapping process is represented as:

[0023] in, and These are the weight matrix and bias vector of the fully connected prediction layer, respectively.

[0024] Furthermore, the method also includes: During the training phase, a loss function is constructed based on the error between the predicted and actual measured values ​​of surface roughness parameters, and the model parameters are updated accordingly. After training, this loss function is used to predict the surface roughness corresponding to the new process combination. Specifically, the total loss function is constructed by weighting the mean square error of the target prediction results with the attention regularization term. , To use the main loss function of mean squared error, The attention standard deviation regularization weight coefficient is used. The standard deviation regularization term for the attention mask is represented using the following logical expression:

[0025] in, This represents the total number of decision steps in the coding layer. Total batch size; The dimension of the input feature; It is the first In the decision-making step, the first The first sample The attention weights corresponding to each feature; Indicates the first In the decision-making step, the first The mean attention of each sample. ; Backpropagation is performed based on the total loss to calculate the gradient of the loss function with respect to all learnable parameters, and the model parameters are updated using the optimizer. After training, the model parameters are fixed, and the new process parameters are input into the trained model to output the corresponding predicted values ​​of the surface roughness of the cylinder bore honing.

[0026] An electronic device includes a memory and a processor, the memory being used to store a program that supports the processor in executing the above-described cylinder bore honing roughness prediction method, the processor being configured to execute the program stored in the memory.

[0027] A storage medium storing a computer program, which, when executed by a processor, performs the steps of the above-described cylinder bore honing roughness prediction method.

[0028] The advantages of this invention are: (1) In the model initialization layer, the present invention introduces a deep cross network (DCN module) and uses the vector outer product operation to construct high-order combined features layer by layer, which explicitly captures the nonlinear physical coupling effect between process parameters such as honing pressure and reciprocating speed, overcoming the limitation that traditional stacked fully connected layers can only implicitly combine features; and enhances the diversity of statistical noise through the random partitioning strategy of virtual sub-batches, which improves the generalization ability of the model while reducing memory overhead, effectively overcoming the problem of weak nonlinear fitting ability of traditional methods, and greatly improving the accuracy of model prediction under complex working conditions.

[0029] (2) This invention avoids the excessive sparsity problem caused by the Sparsemax hard threshold function used in traditional TabNet by adopting a multi-step recursive attention mechanism and an improved continuous feature selection mask generation strategy. This invention replaces the Sparsemax function used in traditional TabNet with Softmax, effectively avoiding the loss of key information caused by excessive feature sparsity; more importantly, the experimental depth of this invention reveals that if the traditional Sparsemax function and the total loss function constructed in this invention are directly superimposed, it will cause a serious “feature collapse” phenomenon, resulting in a significant degradation of model performance. This inversely proves the non-obviousness and absolute necessity of using Softmax for replacement. This invention uses Softmax to generate continuous real number form feature selection masks, ensuring that the attention distribution is smooth and differentiable and does not produce forced sparsity; at the same time, it uses the prior scale to dynamically accumulate and decay with the decision step, so that the attention of features that have been selected at high frequency will naturally decrease in subsequent steps, forcing the model to explore new key parameters and forming a progressive and diversified decision path. This mechanism ensures that the model focuses on core parameters while avoiding the risk of missing key features due to excessive sparsity. Furthermore, the mask output at each step can clearly map the core process parameters, forming a transparent decision path. This reduces the risk of missing key features while ensuring focusing capability, and facilitates reverse guidance for process optimization.

[0030] (3) This invention effectively suppresses measurement noise through systematic coefficient of variation analysis and cross-sectional consistency verification, and at the model end, it takes into account both shared parameters and independent parameter structures in conjunction with an early stopping strategy to ensure that the model still maintains good generalization ability and robustness under small sample conditions.

[0031] Furthermore, by constructing an end-to-end predictive architecture, this invention supports the direct output of prediction results for six key roughness indicators from new process settings. The entire process can be automated, effectively shortening the trial production cycle and reducing production costs, thus meeting the needs of closed-loop quality control in industrial settings. Attached Figure Description

[0032] Figure 1 This is a flowchart of a cylinder bore honing roughness prediction method according to Embodiment 1 of the present invention.

[0033] Figure 2 This is an overall architecture diagram of an improved TabNet prediction model according to Embodiment 1 of the present invention; Figure 3 This is an overall architecture diagram of the feature converter in Embodiment 1 of the present invention; Figure 4 This is an overall architecture diagram of the attention converter in Embodiment 1 of the present invention; Figure 5 This is the first embodiment of the present invention, which uses the true and predicted values ​​based on an independent test set to determine the maximum contour height. A scatter plot comparing the evaluation indicators; Figure 6 This is the first embodiment of the present invention, which uses the peak height based on the true and predicted values ​​of an independent test set. A scatter plot comparing the evaluation indicators; Figure 7 This is the first embodiment of the present invention, which uses the true and predicted values ​​of an independent test set with core roughness depth. A scatter plot comparing the evaluation indicators; Figure 8 This is the valley depth based on the true and predicted values ​​of an independent test set in Embodiment 1 of the present invention. A scatter plot comparing the evaluation indicators; Figure 9 This is the first embodiment of the present invention, which uses the true value and predicted value based on an independent test set in a first material ratio. A scatter plot comparing the evaluation indicators; Figure 10 This is the first embodiment of the present invention, which uses the true value and predicted value based on an independent test set as a second material ratio. A scatter plot comparing the evaluation indicators. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments: Example 1 like Figure 2 As shown, this embodiment provides a prediction model based on an improved TabNet, including an initialization layer, an encoding layer, and a fully connected prediction layer. Figure 2 The initialization layer includes batch normalization modules connected in sequence (FC module); wherein, the initialization layer includes batch normalization modules (FC module) Figure 2 (BN module), Deep cross-network module ( Figure 2 The DCN module, the Feature Transformer module, and the data segmentation module ( Figure 2 The coding layer includes a split module; the coding layer comprises several decision units, each of which includes attention transformers connected in sequence. Figure 2The system consists of an Attentive Transformer module, a Masked module, a Feature Transformer, a Data Segmentation module, and a ReLU activation function layer.

[0036] like Figure 1 As shown, based on the overall structure of the above prediction model, this embodiment provides a method for predicting the roughness of cylinder bore honing based on multi-order feature interaction and continuous attention, including the following steps: S1. Obtain the surface roughness parameters of the cylinder bore honing process and perform preprocessing.

[0037] First, with the reciprocating speed of the oscillator , fine honing speed Jingheng expansion force Honing feed rate Platform honing reciprocating speed Platform honing speed Honing pressure and number of processing times Using process parameters as input features, 64 sets of orthogonal honing experiments were designed and conducted. After each set of experiments, roughness measurements were performed on three sections along the cylinder bore axis: upper, middle, and lower. Four measurement points were evenly selected at different azimuth positions along the circumference of each section, resulting in a total of 12 measurement locations to ensure spatial representativeness of the data. A high-precision roughness measuring instrument was used to obtain the raw measurement values ​​of six key surface morphology indicators, including at least: maximum profile height. Peak height Core roughness depth Valley depth First material ratio Second material ratio .

[0038] Secondly, box plots were used to identify and remove outliers from the original measurement data. Specifically, for the 12 measured values ​​of the same roughness index in each group of experiments, they were arranged in ascending order, and the first quartile was determined according to a preset quartile calculation rule. and the third and fourth quartiles Among them, the first quartile The third quartile is the value at the 25th percentile of the sorted measured value. The upper quartile is the value corresponding to the 75th percentile of the sorted measurement.

[0039] Based on the first quartile and the third and fourth quartiles Determine the threshold range for outlier detection. for:

[0040] in, and These represent the lower and upper limits of the outlier detection threshold interval, respectively. For the interquartile range, less than or greater than The measured values ​​were identified as outliers and removed from the original measurement data of the corresponding group.

[0041] Next, after removing outliers, the coefficient of variation (CV) of the remaining valid measurements in each group is calculated to assess within-group stability. The coefficient of variation is expressed using the following logic:

[0042] in, The standard deviation of the valid measurements in the same group The mean; when When the stability of the measurements within the group was considered good, it met the stability requirements; when At that time, it was considered that there was some measurement fluctuation within the group, but it was still within an acceptable range; when At that time, it was considered that the measurement fluctuations within the group were relatively large, and the potential causes needed to be analyzed in conjunction with the process. Further analysis was conducted on the machining consistency between the upper, middle, and lower sections of the cylinder bore axial direction, and the absolute differences between the sections were calculated for each. Relative proportion to the cross section This can be represented using the following logic:

[0043]

[0044] in, , , These represent the mean roughness values ​​of the upper, middle, and lower sections, respectively. This is the average of the three. When When the three sections are considered to have good consistency, the section consistency requirement is met; when At that time, it was considered that there were certain differences between the cross-sections, but overall they were within an acceptable range; when At that time, it was considered that the cross-sectional differences were large, and further analysis was needed on their relationship with processing pressure, number of processing operations, and local contour changes.

[0045] Finally, considering the outlier removal results, intra-group CV distribution, and cross-sectional consistency, and after confirming that the data has good spatial stability and repeatability, the arithmetic mean of the effective measured values ​​of the corresponding indicators in each group of experiments is taken as the final roughness characterization value under this process combination.

[0046] In this embodiment, the reciprocating speed of the honing is used. , fine honing speed Jingheng expansion force Honing feed rate Platform honing reciprocating speed Platform honing speed Honing pressure and number of processing times A total of 8 process parameters constitute the input feature vector. , with the maximum height of the outline Peak height Core roughness depth Valley depth First material ratio Second material ratio A total of 6 surface morphology indicators constitute the target output vector. .

[0047] After the above preprocessing, a total of 64 sample pairs were formed in each group of experiments. , ,in For the first The 8-dimensional input feature vector of the group experiment, For the first The preprocessed 6-dimensional target output vector of the group of experiments. All samples were divided into training and test datasets in an 8:2 ratio. The training dataset was used for model parameter learning, and the test dataset was used to evaluate the model's predictive performance. The partitioning was performed stratified randomly according to the experiment number to ensure that the training and test sets covered different combinations of process parameters and to avoid evaluation distortion caused by data distribution bias.

[0048] Furthermore, to ensure the fairness and reproducibility of performance comparisons, this embodiment employs a rigorous evaluation process for both the comparison and ablation experiments: First, the 64 orthogonal honing experiments were stratified and randomly divided in an 8:2 ratio, resulting in 51 training data sets and 13 completely independent and unknown test sets. During the training phase, all models underwent hyperparameter optimization using a 5-fold cross-validation strategy within the 51 training data sets to mitigate the risk of overfitting due to small sample sizes. The final performance evaluation metrics (such as MAPE, MAE, RMSE, and...) were... The calculations are performed uniformly on 13 independent test sets to ensure the absolute objectivity and true generalization ability of the model evaluation results.

[0049] Specifically, to quantify the predictive performance of the model on the independent test set, the MAPE (mean absolute percentage error), MAE (mean absolute error), RMSE (root mean square error), and... The formulas for calculating the (determination coefficient) are defined as follows: Mean Absolute Percentage Error (MAPE):

[0050] Mean Absolute Error (MAE):

[0051] Root Mean Square Error (RMSE):

[0052] Coefficient of determination ( ):

[0053] Where n is the total number of samples in the test set; For the first The actual measured true value of the cylinder bore roughness parameter of each sample; The model predicts the first Roughness parameter values ​​for each sample; It is the arithmetic mean of the actual measured values ​​in the test set.

[0054] Among them, MAPE, MAE and RMSE are error measurement indicators. The smaller the value of these three, the smaller the deviation between the model prediction value and the actual value, and the higher the prediction accuracy. It belongs to the goodness-of-fit index, and its value ranges from 0 to 1. The larger the value (the closer to 1), the stronger the model's ability to interpret the real data and the better the fit.

[0055] S2. The input features are fed into the initialization layer for multi-stage feature preprocessing and feature enhancement, generating the initial backbone feature vector and attention guidance signal. Specifically, the functional modules within the initialization layer and the multi-stage feature preprocessing and feature enhancement process in S2 are further explained: S21. Perform batch normalization on the input features and introduce virtual sub-batch normalization (Ghost Batch Normalization, Ghost BN) to accelerate the training process. Divide the input batch data used for training into several virtual sub-batches (ghost batches), so that each virtual sub-batch independently performs mean and variance statistics, and outputs the normalized input features.

[0056] Specifically, let the total batch size be... Divide it evenly into There are 1 sub-batch, each sub-batch being 1 unit in size. For the first Samples in a virtual sub-batch Normalized samples Using the following logical representation:

[0057] in, For the first Input features of each sample; For the first The mean of each virtual sub-batch ; For the first The variance of each virtual sub-batch ε is a smoothing term used to prevent division by zero; in this embodiment, it is preferably taken as ε. .

[0058] In the initialization layer, the Ghost BN module can significantly reduce memory usage and computational overhead without changing the model's expressive power. At the same time, the random partitioning strategy of virtual subbatches enhances the diversity of statistical noise, effectively improving the model's generalization ability and reducing dependence on the statistical characteristics of specific batches.

[0059] Furthermore, the samples in each sub-batch after virtual sub-batch normalization processing The samples are re-merged into complete, batch-normalized input features according to their order in the original batch. , As a complete batch of data, it is used to perform layer-by-layer construction of higher-order combined features.

[0060] S22. Construct higher-order combined features layer by layer through vector outer product operations: Step S22 introduces a Deep & Cross Network (DCN) module, which constructs second-order and higher-order feature combinations layer by layer by building feature outer products to explicitly model the high-order nonlinear interaction relationships between input process parameters, thereby capturing the coupling effect between process parameters; specifically, in the... The layer uses the following logic to perform cross-calculation to construct higher-order combined features. :

[0061] in, The input features are after batch normalization. This represents the vector outer product operation. For the first The current layer state vector is used to capture the combination relationship of second-order and higher features; and These are the learnable weight matrix and bias term for the corresponding layer, respectively.

[0062] In this embodiment, a DCN module is introduced into the initialization layer, enabling the deep cross-network to gradually accumulate high-order combination features, significantly enhancing the ability to perceive key interactions such as oscillation pressure and reciprocating speed.

[0063] S23. Perform multi-layer nonlinear mapping on the enhanced high-order combined features, and divide the first output feature after nonlinear mapping into two sub-vectors along the feature dimension, which are used as the initial backbone feature vector and the initial attention guidance signal, respectively: Furthermore, the enhanced features are input into the first feature transformer (Feature Transformer module) for nonlinear transformation, and the initial backbone features are separated by the Split module. With initial attention guidance signals Among them, the initial backbone features Entering the coding layer to participate in sparse selection, while the initial attention guide signal... Used to drive the next feature focusing mechanism.

[0064] Specifically, after completing the various preprocessing operations of the initialization layer, the input features are transformed into high-quality initial representations. The data is fed into the first Feature Transformer module for preliminary nonlinear transformation to obtain the output features. Subsequently, it is divided into two sub-vectors along the feature dimension using a Split operation, represented by the following logic:

[0065] in, The initial backbone feature vector is used as the input for the first step of the encoding layer, and is then used to participate in subsequent sparse feature selection and high-level semantic learning. It is the set of real numbers; The initial attention guidance signal serves as input to the Attentive Transformer module in the next decision step, used to generate the feature selection mask for the next step. ; The operation function for segmentation along the feature dimension typically employs an equal-segmentation strategy, but asymmetric segmentation ratios can also be set according to task requirements. In this embodiment, the Split mechanism decouples the feature flow from the attention flow, enabling the model to dynamically update its selection strategy at each step based on historical attention states, forming a progressive and interpretable multi-step decision-making process.

[0066] In this embodiment, the Feature Transformer module consists of multiple feature blocks, including multiple shared feature blocks and multiple step feature blocks, to achieve efficient and flexible nonlinear feature transformation; wherein, the shared feature blocks reuse parameters across different decision steps, while the step feature blocks are configured with independent parameters and combined with a residual connection structure with scaling factors.

[0067] In this embodiment, the Feature Transformer module, as the core feature processor, is responsible for performing multi-layer nonlinear mapping on the input features and focusing on key process parameters using a dynamic masking mechanism. For example... Figure 2 As shown, in this embodiment, Feature Transformer modules are provided at different levels. In the initialization layer, the output features of the first Feature Transformer module of the model are... Represented as:

[0068] in, These are higher-order combinatorial features after passing through the DCN module. b is the batch size. For feature dimensions; Choose a mask for the first feature, and take... The algorithm uses a broadcast mechanism to multiply the input features element-wise to ensure that all features are involved in the initial computation. This indicates element-wise multiplication; The first feature processing function consists of several fully connected layers, a batch normalization module, a GLU activation function, and residual connections, and is used to extract abstract semantic features. This is the high-order feature tensor output by the current Feature Transformer module. The output dimension is usually the same as the input dimension or adjusted by projection.

[0069] In a preferred embodiment, the feature blocks in the Feature Transformer module are structurally divided into the following two categories: Shared blocks share network parameters across all decision steps, meaning the same set of weights is repeatedly applied across multiple time steps. This design significantly reduces the total number of model parameters, lowers the risk of overfitting, and improves generalization ability and computational efficiency. It is suitable for learning common basic feature patterns across steps.

[0070] Step Blocks allow for independent parameter configuration for each decision step, enabling the network to dynamically adjust its feature extraction behavior based on the current context and enhancing the flexibility of modeling local nonlinear relationships under specific operating conditions.

[0071] In a preferred embodiment, to alleviate the gradient vanishing / exploding problem in deep networks, residual connections are introduced in all shared feature blocks except the first layer, and the residual paths are multiplied by a scaling factor. To stabilize variance propagation.

[0072] like Figure 3 As shown, the Feature Transformer module consists of several fully connected layers, a batch normalization module, a GLU activation function, and residual connections. The Feature Transformer module extracts features through fully connected layers, improves model stability and convergence speed through the batch normalization module, and uses GLU activation function layers for nonlinear transformation and feature selection, which helps the model learn complex feature representations of the input data. Internally, the Feature Transformer uses a Gated Linear Unit (GLU) as the nonlinear activation function to enhance feature representation and control information flow. The GLU determines which parts of the input features should be preserved or suppressed through a nonlinear gating mechanism. Its calculation process is as follows:

[0073] in, Represents higher-order combination features of the input The result of the linear transformation For learnable weight matrix, For bias terms, It is a high-order combination feature vector The result after linear transformation; The Sigmoid activation function is used to generate gate weights, and its value range is... ; For linear path output, i.e. ; This is element-wise multiplication, meaning that corresponding positions are multiplied.

[0074] The Feature Transformer module structure allows the network to adaptively adjust the information gain of each feature path, improving nonlinear modeling capabilities while mitigating the gradient vanishing problem. It is particularly suitable for capturing complex nonlinear relationships in high-dimensional tabular data.

[0075] S3. Input the initial attention guidance signal into the encoding layer, and perform sparse feature selection and abstract representation learning through a multi-step recursive decision mechanism to output a series of encoded features.

[0076] The encoding layer adopts a decision tree-like stepwise focusing architecture, consisting of... The code consists of several identical decision units connected in series. Each decision unit comprises an Attention Transformer, a Masked module, a Feature Transformer, a Split module, and a ReLU activation function layer, all connected in sequence. These functional modules serve as the core components, jointly enabling the dynamic selection and extraction of key process parameters. Specifically, the detailed process of each functional module within the encoding layer and the multi-step recursive decision-making mechanism in S3 is further explained below: S31. In each decision step, the attention transformer receives the attention guidance signal output from the previous step, and after being processed by the linear mapping and batch normalization module in the fully connected layer, it generates a feature selection mask in continuous real number form by combining the learnable historical attention intensity adjustment vector.

[0077] like Figure 4 As shown, the attention transformer includes a fully connected layer (FC), a batch normalization module (BN), and a softmax activation function layer connected in sequence, and introduces a prior scale vector. In this embodiment, the attention transformer abandons the Sparsemax function used in the traditional TabNet model and adopts a softmax attention mechanism to generate feature selection masks. .

[0078] In this embodiment, replacing the Sparsemax function with the Softmax function is not a conventional replacement in this field. Its technical necessity lies in the fact that the Softmax function effectively avoids information loss caused by excessive feature sparsity, alleviating the problem of missing key features. More importantly, subsequent experiments in this embodiment reveal that directly superimposing the traditional Sparsemax with the loss function module constructed in this embodiment will cause a severe "feature collapse" phenomenon and lead to a serious degradation of model performance. This conversely proves that using Softmax replacement is a necessary condition for ensuring the effectiveness of the overall model architecture. Specifically, the first... Step Feature Selection Mask Using the following logical representation:

[0079] in, This serves as an attention guidance signal for the previous step. It is a mapping function consisting of a fully connected layer and a batch normalization module, used to convert the attention signal into the original feature space dimension; This is the historical attention intensity adjustment vector that can be learned in the previous step, also known as the prior scale vector, which is used to adjust the historical attention intensity of each feature channel. This is element-wise multiplication; The standardized exponential function is defined as:

[0080] Where z is the input vector of the Softmax function, specifically referring to the weighted mapping feature in the above equation in this step, i.e. ; This represents the first input vector. Each element. This represents the dimension of the mask vector, which is the same as the dimension of the output features of the feature transformer in step S23. Maintain complete consistency to ensure element-wise multiplication with features is possible subsequently. The output of the Softmax function represents the... Each feature corresponds to a soft attention weight. This function maps the input to a normalized weight distribution that satisfies non-negativity and sums to 1. This approach not only achieves smooth allocation of feature weights but also ensures the continuous differentiability of the network during backpropagation.

[0081] S32. The historical attention intensity adjustment vector is dynamically updated with each decision step, and the attention priority of the corresponding channel is gradually reduced based on the feature selection mask generated in the previous stage.

[0082] In this embodiment, the historical attention intensity adjustment vector It updates dynamically with each decision step, and its update rules are as follows:

[0083] in: For the first Feature selection mask generated in each decision step ; As a relaxation factor, , These are adjustable hyperparameters used to control the probability of historical features being selected again; when any feature has been frequently selected in the past, its corresponding channel's... The value gradually decreases, thereby reducing the possibility of repeated activation in subsequent steps; this mechanism effectively improves the diversity of feature selection and avoids the model getting trapped in local optima.

[0084] S33. The masking module performs element-wise multiplication of the generated feature selection mask with the input features received by the initialization layer, and outputs the masked features to achieve selective suppression of the input features.

[0085] This embodiment designs a mask module to implement attention-based feature selection operations and construct a tree-like decision manifold. Specifically, the mask module receives feature selection masks from the attention transformer. and the normalized features output in step S21 Perform a mask weighting operation to generate the masked features. This can be represented using the following logic:

[0086] in, This indicates element-wise multiplication. This represents the input features after batch normalization. The mask module supports continuous value masks, i.e. The elements in the algorithm take values ​​that are real numbers in the interval (0,1), rather than strictly binary (0 / 1). After element-wise multiplication, key features with larger weights are retained, while irrelevant features with smaller weights are smoothly suppressed. This soft selection mechanism makes the entire feature selection process end-to-end differentiable, fully supports backpropagation optimization, and simulates the function of selecting feature branches based on conditions in a decision tree.

[0087] S34. Input the masked features into the feature transformer of the current step for nonlinear transformation, and output the main features of the current step and the new attention guidance signal. After repeating the above process cyclically according to the set total number of steps, all the main features output will be aggregated for target prediction.

[0088] In this embodiment, the features after masking are... The Feature Transformer module takes the current step as input, performs nonlinear transformation and feature abstraction, and then outputs the features. Then, after the Split operation, the main features of the current step are output. and new attention-guiding signals This is for use in the next iteration. The output features of the Feature Transformer module are... Represented as:

[0089] in, It is the input feature vector after batch normalization, where b is the batch size. For feature dimensions; For the first Feature selection mask for each decision step; This represents the element-wise multiplication operation, used to sparsify the input features and suppress the influence of unimportant parameters. ; For the first The feature processing function in each decision step consists of several fully connected layers, a batch normalization module, a GLU activation function, and residual connections. It is used to extract abstract semantic features. The calculation process of GLU is the same as in S23, except that the input of GLU is not a high-order combination of features. Instead ; This is the high-order feature tensor output by the current Feature Transformer module. The output dimension is usually the same as the input dimension or adjusted by projection.

[0090] The entire encoding process is repeated This forms a progressive, sparse, multi-step decision-making process, ultimately outputting all the core features. It is aggregated and used as input to the fully connected prediction layer to generate target surface roughness parameters.

[0091] It should be noted that when the encoding layer passes information to the fully connected prediction layer, it only aggregates the backbone features separated from each decision step as decision features, without including the attention guidance signals generated by the split module after each decision. The attention guidance signals generated by each decision only flow within the encoding layer and are specifically used to drive the next attention transformation module to generate the feature selection mask, thereby decoupling the information flow of feature prediction and feature selection.

[0092] S4. Input the decision representation output from the encoding layer into the fully connected prediction layer, map it to the target space, and generate the predicted surface roughness value for cylinder bore honing. Specifically, S4 includes the following: S41. Perform nonlinear activation and aggregation on the decision features output by each decision step of the encoding layer to generate global decision features for target mapping: In this embodiment, after the input process parameters are processed by the initialization layer and the encoding layer, corresponding local decision features are generated in multiple consecutive decision steps. It is important to note that when the encoding layer transmits information to the fully connected layer, it only aggregates the backbone features (i.e., decision features) separated from each decision step, and does not include the attention guidance signal generated for each decision. The attention guidance signal generated for each decision flows only within the encoding layer and is specifically used to drive the next attention transformer to generate a feature selection mask, thereby decoupling the feature prediction and feature selection information flows. Assume that the encoding layer contains... The first decision-making step, the... The decision characteristics separated from each decision step are represented as follows: ,in The aforementioned The core features are extracted by nonlinear mapping through the Feature Transformer and feature segmentation in the current step, which contain information on the combination of key process parameters that the current step focuses on. The total number of decision steps is denoted as , and the adjustable hyperparameter is denoted as 3 to 6.

[0093] To integrate the feature extraction results of all decision-making steps and ensure that the local decisions of multiple steps can be effectively fused into a single global feature, the network employs an accumulation operation with a non-linear activation function to aggregate the local decision features output by each decision step, thereby obtaining the global decision feature output by the encoding layer. The polymerization process is represented as follows:

[0094] in, ∑ is the rectified linear unit activation function, used to enhance the nonlinearity and sparsity of feature representation; ∑ is the accumulation operation; These are the global decision features generated after aggregation. It integrates high-level nonlinear features extracted from different decision-making steps to comprehensively characterize the complex mapping relationship between input process parameters and surface roughness indices.

[0095] S42. The aggregated global decision features are mapped to the target space through a fully connected prediction layer to generate the final surface roughness prediction value: The global decision features obtained through multi-step accumulation and aggregation The uniform input is fed to the fully connected prediction layer (FC) at the end of the encoding layer, and linearly projected onto the target parameter space to obtain the final surface roughness prediction vector of the model. The mapping process is represented as follows:

[0096] in, and These are the weight matrix and bias vector of the fully connected prediction layer, respectively, which are jointly optimized during the model training phase using the backpropagation algorithm.

[0097] The dimension of the target variable to be predicted is Then the predicted vector output by the model is represented as Preferably, when the metric to be predicted is the maximum height of the profile... Peak height Core roughness depth Valley depth First material ratio Second material ratio When the prediction vector is specified, it is represented as ], m=6.

[0098] S5. Construct a total loss function based on the prediction loss and attention regularization term, update the model parameters through backpropagation, and perform surface roughness prediction after training. Specifically, the training phase includes the following: S51. During the training phase, Mean Squared Error (MSE) is used as the main loss function to measure the deviation between the model's prediction results and the true labels. The main loss function is defined as follows:

[0099] in, This is the overall loss value, a scalar, representing the average prediction error on the current batch of samples; The number of samples in the current training batch (batch size); For the first The vector of actual measured values ​​for each sample, corresponding to Surface roughness index ( The data were collected by a German Mahr roughness measuring instrument and obtained after outlier removal and mean value processing. For the model to the first A surface roughness prediction vector is generated for each sample.

[0100] S52. To further enhance the model's feature diversity and representational power, an attention standard deviation regularization term is introduced into the total loss function. This term constrains the model to focus on differentiated information from different feature channels during feature selection, preventing the attention distribution from becoming homogeneous. The total loss function is represented by a weighted sum of the main loss function and the attention standard deviation regularization term (ASDR). The diversity weight coefficients are determined through an optimization algorithm. This can be represented using the following logic:

[0101] in, For the ultimate optimization goal, used for parameter updates; is the attention standard deviation regularization weight coefficient, and is an adjustable hyperparameter used to control the dispersion of attention across different feature dimensions. In this embodiment, the hyperparameter search space is defined and the Hyperopt algorithm is used to optimize the hyperparameters.

[0102] In a preferred embodiment, hyperparameters , The standard deviation regularization term of the attention mask promotes a differentiated distribution of attention weights for each feature channel during feature selection, enhancing the model's ability to capture multi-feature patterns. Its calculation formula is as follows:

[0103] in, This represents the total number of decision steps in the coding layer. The dimension of the input feature; Total batch size; It is the first In the decision-making step, the first The first sample The attention weights corresponding to each feature; Indicates the first In the decision-making step, the first The mean attention of each sample.

[0104] S53. Based on the above loss function, backpropagation is performed using an automatic differentiation framework to calculate the gradient of the loss with respect to all learnable parameters, and the model parameters are updated using the Adam optimizer:

[0105] in, , These are the sets of all trainable parameters of the model at step t+1 and step t, respectively, including trainable parameters such as DCN weights, Feature Transformer module, and Attentive Transformer module network weights. For training steps; The learning rate typically ranges from 1 to 10. It can be dynamically adjusted through scheduling strategies; For the loss function with respect to the parameters The gradient vector.

[0106] Furthermore, an early stopping mechanism is implemented during training to prevent overfitting: the loss on the validation set is monitored, and if continuous loss changes occur... In-wheel verification loss did not decrease beyond the preset threshold. If the training is terminated prematurely, then the training is terminated early; preferably, , Pick .

[0107] S54. After completing the training, fix the model parameters, input the new process parameter combination into the trained model, and after the initialization layer enhancement, encoding layer sparse selection, multi-step feature aggregation and target mapping, output the corresponding cylinder bore honing surface roughness prediction value, so as to achieve efficient and accurate prediction of the processing quality under unseen working conditions.

[0108] Furthermore, to verify the effectiveness of the cylinder bore honing roughness prediction method described in this embodiment, the honing process of a certain type of four-cylinder cast iron engine block is used as an example to illustrate the workflow of the prediction method proposed in this embodiment in detail: (1) Hyperparameter configuration and experimental design: In this embodiment, to ensure the model possesses optimal modeling capabilities, the Hyperopt algorithm is used to perform a global search and optimization of key hyperparameters. The final set of optimal hyperparameters is as follows: feature hiding dimension is set to 24, attention transformation dimension to 16; decision steps are 4, which effectively extracts multi-stage features while avoiding redundant computation; relaxation factor is set to approximately 1.26 to dynamically adjust the attention decay intensity of historical features; and the sparsity weight coefficient is approximately... To encourage the attention mechanism to focus on core process parameters, a learning rate of 0.048 was set, coupled with the Adam optimizer to accelerate convergence. During training, the batch size was 32, and the virtual sub-batch size was also 16 to improve normalization stability. The maximum number of training epochs was set to 1000, and an early stopping mechanism was enabled, terminating training prematurely when the validation loss did not show a significant decrease within 80 consecutive epochs to prevent overfitting. Furthermore, the deep cross-network layer was set to 1 layer to ensure high-order interaction modeling while controlling model complexity.

[0109] (2) Data acquisition and preprocessing: This embodiment conducted 64 orthogonal honing experiments, covering typical process windows. Each experiment set different combinations of input parameters, including eight key process variables: fine honing reciprocating speed, fine honing rotation speed, fine honing expansion force, fine honing feed speed, platform honing reciprocating speed, platform honing rotation speed, honing pressure, and number of processing cycles. The corresponding output response consists of six surface roughness indicators: maximum profile height. Peak height Core roughness depth Valley depth First material ratio Second material ratio .

[0110] After each set of experiments, roughness measurements were performed on three cross-sections along the cylinder bore axis: upper, middle, and lower. Four measurement points were evenly selected at different locations along the circumference of each cross-section, for a total of 12 measurement positions to ensure spatial representativeness. Data acquisition was completed using a German Marco precision roughness measuring instrument, with the equipment parameters set to a vertical magnification of 20.0. Horizontal magnification 2.5 , assessment length =15mm, conforming to ISO 4287 / 4288 international standards.

[0111] To address the random errors and local fluctuations in the original measurement data, outliers were first identified and removed using box plots. Specifically, the lower quartiles were calculated based on 12 measurements of the same roughness index in each experimental group. and upper quartiles Define the normal range as Measurements outside this range were considered outliers and deleted. A total of 206 outliers were identified and removed, representing 4.47% of all 4608 measurements, indicating that the overall quality of the original data was good.

[0112] After outlier removal, further intragroup stability analysis was conducted. The coefficient of variation of valid measurements for each group was calculated to assess the dispersion among multiple measurement points within the same cylinder bore. The results showed that... , The average coefficient of variation for these indicators is less than 10%, demonstrating good repeatability; while Because it is sensitive to local contour peaks, the average coefficient of variation is relatively high, but still within an acceptable range.

[0113] Simultaneously, cross-sectional consistency analysis was performed, calculating the maximum difference (absolute difference) in the mean roughness values ​​of the upper, middle, and lower cross-sections and its ratio relative to the average of the three (relative proportion). For example, The mean relative difference of the cross sections of the index was only 1.58%, and all experimental groups were less than 10%, indicating that their axial distribution was highly consistent. , , The relative differences in cross sections are also generally low, indicating that the processing process has good axial uniformity.

[0114] Based on the above analysis, and assuming that the data has good integrity, stability and consistency, the arithmetic mean of the effective measured values ​​of each index in each group of experiments is taken as the final roughness characterization value under the process combination, which is used to construct a high-quality training sample set.

[0115] (3) Feature initialization and enhancement: The standardized 8-dimensional process parameters are input into the model initialization layer. First, batch normalization is performed, and Ghost Batch Normalization is introduced to divide the current training batch into multiple virtual sub-batches. Each sub-batch independently performs mean and variance estimation, thereby enhancing the statistical diversity during training and improving the model's generalization ability.

[0116] Next, the features are fed into the Deep & Cross Network module to explicitly model high-order nonlinear interactions. For example, the product of "honing force" and "reciprocating speed" is constructed layer by layer, effectively capturing the physical coupling effect of the two influencing the cutting depth.

[0117] After feature enhancement, the features are fed into the first Feature Transformer module for initial nonlinear transformation. Then, a Split operation divides the output features into two equal parts along the dimensional direction: one part serves as the backbone feature vector, participating in the sparse selection of subsequent encoding layers; the other part serves as the attention guidance signal, driving the feature focusing mechanism in the next decision step. This mechanism decouples the feature flow from the attention flow, supporting a progressive multi-step inference process.

[0118] (4) Sparse feature selection in the coding layer: The encoding layer employs a decision tree-like progressive focusing architecture, consisting of four identical decision units strung together. Each decision unit contains an attention transformer, a masking module, and a feature transformer.

[0119] In the first step, the attention transformer receives an initial attention guidance signal and, combined with a learnable prior scale vector, generates the first feature selection mask. Visual analysis shows that this mask assigns the highest weights to the channels corresponding to "honing pressure" and "fine honing speed," indicating that the model prioritizes the core parameters that dominate the material removal process.

[0120] In subsequent steps, the prior scale vector dynamically decays its reactivation probability based on previously selected features, enabling the model to proactively explore new important parameters. For example, in the second step, attention shifts to "platform honing speed" and "number of processing cycles," reflecting the focus on texture trimming and cumulative effects during the platform honing stage; the third step focuses on the synergistic effect of "fine honing feed speed" and "platform honing reciprocating speed."

[0121] The entire coding process forms a clear multi-step decision-making path, with each step focusing on a specific set of key parameters, which greatly enhances its interpretability and enables engineers to intuitively understand the model's reasoning logic.

[0122] (5) Multi-step feature aggregation and target mapping: The local decision features extracted from multiple consecutive decision-making steps in the encoding layer are nonlinearly activated and fused into global decision features through accumulation, thereby deeply representing the key information of the input process parameters. Subsequently, this global decision feature is input into the fully connected prediction layer, and linear projection is performed using weights and bias parameters to accurately map it to the target parameter space, ultimately generating a surface roughness prediction vector containing multiple indicators such as maximum profile height and peak height.

[0123] By nonlinearly activating and accumulating local features, key information from each decision-making stage is effectively integrated to construct a global feature representation that comprehensively reflects the process mapping relationship. This mechanism, combined with the target space projection of the fully connected layer, significantly enhances the model's ability to nonlinearly express complex process parameters, thereby achieving efficient and high-precision joint prediction of multidimensional processing quality indicators.

[0124] (6) Model training and prediction: During the training phase, the deviation between the predicted and actual measured values ​​is used as the main loss function to measure the overall fit of the model. Simultaneously, to enhance the sparsity of the attention mechanism, a regularization penalty term is introduced into the total loss, encouraging the model to focus on only a few key input features at each step, thus avoiding attention diversion.

[0125] The gradient is calculated using the backpropagation algorithm, and the Adam optimizer is used to update all trainable parameters in the model, including the fully connected layer parameters in the DCN weights, the attention mapping matrix, and the weights of the feature transformation network.

[0126] During training, the loss on the validation set is continuously monitored. If no significant performance improvement is observed for 80 consecutive rounds, an early stopping mechanism is triggered, training is terminated early, and the model state at which the validation loss is lowest is retained to prevent overfitting.

[0127] After training is complete, the model is deployed on the edge computing terminal of the production line. After new process parameters are input, the system can complete end-to-end inference and output predicted values ​​for six surface roughness indices.

[0128] like Figures 5-10 As shown, in order to intuitively verify the predictive performance of the trained model, Figures 5-10 A scatter plot comparing predicted and true values ​​on 13 independent test sets of a trained prediction model. The vertical axis represents the predicted value output by the model, and the horizontal axis represents the actual measured value obtained by the high-precision roughness measuring instrument. The diagonal dashed line in the figure represents the ideal prediction state (i.e., the case where the predicted value is exactly equal to the true value). Figures 5-10 These represent six key surface roughness indicators: maximum profile height. Peak height Core roughness depth Valley depth First material ratio Second material ratio The prediction effect is shown in the upper left corner of the figure, which represents four quantitative evaluation indicators (R², MAE, RMSE, and MAPE).

[0129] Depend on Figures 5-10 As can be seen, the test data points for the six roughness indices are all closely distributed on both sides of the diagonal dashed line. Among them, the core roughness depth... First material ratio Second material ratio The coefficients of determination (R²) for all indicators exceeded 0.94 (reaching a maximum of 0.9775), and the mean absolute percentage error (MAPE) for each indicator remained at a low level (most were controlled within 7%). This fully demonstrates that, after feature enhancement in the initialization layer and multi-step recursive sparse selection in the encoding layer, the prediction model can accurately capture the complex nonlinear mapping relationship between honing process parameters and the microstructure of the cylinder bore surface, exhibiting excellent prediction accuracy and generalization ability even under unseen operating conditions.

[0130] Furthermore, this embodiment selects several existing deep learning models to predict the roughness of cylinder bore honing, and compares them on an independent test set using mean absolute percentage error (MAPE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) as performance evaluation indicators. The deep learning models include various benchmark models such as ELM, GPR, SVR, RF, XGBoost, CatBoost, MLP, ANN, CNN, ResNet, and TabNet. The performance of each model in predicting the roughness of cylinder bore honing is shown in Table 1 below. It can be seen that the prediction model proposed in this embodiment (Ours) is significantly better than the above benchmark models in terms of mean absolute percentage error, mean absolute error, root mean square error, and coefficient of determination.

[0131] Table 1. Performance comparison results of different benchmark models on independent test sets.

[0132] To verify the effectiveness and synergistic effect of each component in the proposed prediction model, an ablation experiment was designed on an independent test set. Using the standard TabNet as the baseline model, key components were gradually replaced or added / removed, including the attention mechanism (Softmax replacing Sparsemax), the deep cross-network (DCN), and attention standard deviation regularization (ASDR), until the complete structure of the proposed prediction model (corresponding to TabNet+softmax+DCN+ASDR in Table 2) was finally formed. Table 2 below lists the prediction performance under eight different component combinations: Table 2 Ablation experiment results of key components of the prediction model in this embodiment on independent test sets.

[0133] Ablation experiments verified the synergistic advantages of the components of the proposed model in this embodiment: First, replacing Sparsemax with Softmax effectively avoids information loss caused by excessive feature sparsity. The experiments also revealed that the superposition of Sparsemax and ASDR module will cause feature collapse and serious performance degradation, which in turn proves the necessity of replacing Softmax.

[0134] Building upon this foundation, this embodiment proposes a prediction method that perfectly solves the module compatibility problem. The spatial feature adaptation of the DCN module and the regularization of the ASDR module achieve deep synergy on Softmax. Results show that the proposed model not only reduces all prediction errors to the lowest globally, but also demonstrates that the proposed method achieves excellent prediction accuracy while possessing superior robustness and generalization stability, highly meeting the actual prediction needs of complex industrial scenarios.

[0135] Example 2 An apparatus includes a memory and a processor, the memory being used to store a program that supports the processor in executing the cylinder bore honing roughness prediction method of Embodiment 1, the processor being configured to execute the program stored in the memory.

[0136] Example 3 A storage medium storing a computer program, which, when executed by a processor, performs the steps of the cylinder bore honing roughness prediction method in Embodiment 1.

[0137] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein can be considered as a ordered list of executable instructions for implementing logical functions, which can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-based system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0138] For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit a program for use in or in conjunction with an instruction execution system, apparatus, or device. More specific examples of computer-readable media (a non-exhaustive list) include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CD-ROM). Additionally, a computer-readable medium can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optical scanning of the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0139] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0140] In the description of this specification, references to terms such as "in a preferred embodiment," "preferred," "in this embodiment," "specific," or "furthermore," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0141] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the roughness of cylinder bore honing, characterized in that, Includes the following steps: Determine the input features, obtain surface roughness data of cylinder bore honing under different process combinations, and determine the target output vector; The input features are batch normalized in the initialization layer, and the cross product of the normalized features and the current layer state vector is calculated to construct higher-order combined features layer by layer. Multi-layer nonlinear mapping is performed on the higher-order combined features, and the first output feature after nonlinear mapping is divided into an initial backbone feature vector and an initial attention guidance signal along the feature dimension. The initial attention guidance signal is input into the encoding layer, a feature selection mask is generated in each decision step, and a new backbone feature vector and attention guidance signal are generated by combining the input features; the new attention guidance signal is used for the next decision step, and the new backbone feature vector is input into the fully connected prediction layer. The fully connected prediction layer performs nonlinear activation and aggregation on the backbone feature vectors output from multiple decision steps, maps them to the target space, and generates the predicted target output vector.

2. The method for predicting the roughness of cylinder bore honing according to claim 1, characterized in that, The input features include the reciprocating speed of the oscillator. , fine honing speed Jingheng expansion force Honing feed rate Platform honing reciprocating speed Platform honing speed Honing pressure and number of processing times ; The target output vector includes the maximum contour height. Peak height Core roughness depth Valley depth First material ratio Second material ratio ; Orthogonal honing experiments were conducted, and measurement points were arranged at the upper, middle and lower sections along the cylinder bore axis to perform multiple roughness measurements and collect the original surface roughness data. For the measurement results of the same index in each group of experiments, the threshold range for judging outliers for each index is determined by using the interquartile range as the interval. If the value exceeds the range, it is considered an abnormal measurement value and is removed. Calculate the coefficient of variation (CV) of the remaining valid measurements to assess spatial stability, and calculate the cross-sectional difference index between the upper, middle, and lower sections, including the absolute difference between the sections. Relative proportions of cross sections ; The arithmetic mean of the effective measured values ​​that meet the preset requirements of coefficient of variation and cross-sectional difference index in each group of experiments is taken as the final roughness characterization value under the process combination. After the above preprocessing, multiple sample pairs were formed in each group of experiments. ,in For the first The input feature vector of the group experiment, For the first The preprocessed target output vector of the group experiment is used to divide all samples into training dataset and test dataset according to the proportion.

3. The method for predicting the roughness of cylinder bore honing according to claim 1, characterized in that, The process of constructing the higher-order combined features is as follows: In the The layer uses the following logic to perform cross-computation to construct higher-order combined features. : in, The input features are after batch normalization. This represents the vector outer product operation. For the first The current state vector of the layer, and These are the learnable weight matrix and bias term for the corresponding layer, respectively.

4. The method for predicting the roughness of cylinder bore honing according to claim 1, characterized in that, The process of generating the initial backbone feature vector and the initial attention guidance signal is as follows: First, the higher-order combined features are input into the first feature transformer for nonlinear transformation to obtain the output features. ; wherein, the output feature Using the following logical representation: in, It is a high-order combinatorial feature. b is the batch size. The dimension of the input feature. It is the set of real numbers; Choose a mask for the first feature, and take... ; For element-wise multiplication; The first feature processing function consists of several fully connected layers, a batch normalization module, a GLU activation function, and residual connections. The calculation process is as follows: in, It is a high-order combination feature vector The result after linear transformation , For learnable weight matrix, For bias terms; It is the Sigmoid activation function. Output for linear path ; Next, output features The initial backbone features are separated using the Split module. With initial attention guidance signals This can be represented using the following logic: in, This is the initial backbone feature vector. As an initial attention-guiding signal, This is the operation function for segmenting along the feature dimension.

5. The method for predicting the roughness of cylinder bore honing according to claim 1, characterized in that, The process of generating the new backbone feature vector and attention guidance signal is as follows: In each decision-making step, the attention transformer receives the attention guidance signal output from the previous step, sequentially passes it through a fully connected layer for linear mapping and batch normalization, and then combines it with a learnable historical attention intensity adjustment vector. Generate feature selection masks in continuous real number form The attention transformer uses a Softmax attention mechanism to generate a feature selection mask for the th... Step Feature Selection Mask This can be represented using the following logic: in, It is a standardized exponential function; It is a mapping function consisting of a fully connected layer and a batch normalization module; This serves as an attention guidance signal for the previous step. This indicates element-wise multiplication; The historical attention intensity adjustment vector is learned from the previous step. The historical attention intensity adjustment vector is dynamically updated with the number of decision steps, and the attention priority of the corresponding channel is gradually reduced based on the mask selected by the previously generated features. The masking module performs element-wise multiplication of the generated feature selection mask with the batch-normalized input features, and outputs the masked features. ; Features after masking Input the feature transformer of the current step to perform nonlinear transformation and output features Then, after the Split operation, the main features of the current step are output. With new attention-guiding signals ; wherein, the output feature Using the following logical representation: in, The input features are after batch normalization. For the first Feature selection mask for each decision step; For the first The feature processing function in each decision step consists of several fully connected layers, a batch normalization module, a GLU activation function, and residual connections. Based on the set total number of steps After repeating the above process cyclically, all the main features will be output. polymerization.

6. The method for predicting the roughness of cylinder bore honing according to claim 5, characterized in that, The historical attention intensity adjustment vector It updates dynamically with each decision step, and its update rules are as follows: in: For the first Feature selection mask generated in each decision step ; The relaxation factor is an adjustable hyperparameter; when any feature has been frequently selected in the early stages, the corresponding channel's... The value gradually decreases.

7. The method for predicting the roughness of cylinder bore honing according to claim 1, characterized in that, The process of generating the predicted surface roughness value for the cylinder bore honing is as follows: An accumulation operation with a non-linear activation function is used to aggregate the backbone features output from each decision step to obtain the global decision features output from the encoding layer. The polymerization process is represented as: in, Let ∑ be the linear rectified activation function, and let ∑ be the accumulation operation. For the first The main features output by each decision step This represents the total number of decision-making steps. The aggregated global decision features are linearly projected into the target parameter space to obtain the surface roughness prediction vector for cylinder bore honing. The mapping process is represented as: in, and These are the weight matrix and bias vector of the fully connected prediction layer, respectively.

8. The method for predicting the roughness of cylinder bore honing according to claim 1, characterized in that, The method further includes: During the training phase, a loss function is constructed based on the error between the predicted and actual measured values ​​of surface roughness parameters, and the model parameters are updated accordingly. After training, this loss function is used to predict the surface roughness corresponding to the new process combination. Specifically, the total loss function is constructed by weighting the mean square error of the target prediction results with the attention regularization term. , To use the main loss function of mean squared error, The attention standard deviation regularization weight coefficient is used. The standard deviation regularization term for the attention mask is represented using the following logical expression: in, This represents the total number of decision steps in the coding layer. Total batch size; The dimension of the input feature; It is the first In the decision-making step, the first The first sample The attention weights corresponding to each feature; Indicates the first In the decision-making step, the first The mean attention of each sample. ; Backpropagation is performed based on the total loss to calculate the gradient of the loss function with respect to all learnable parameters, and the model parameters are updated using the optimizer. After training, the model parameters are fixed, and the new process parameters are input into the trained model to output the corresponding predicted values ​​of the surface roughness of the cylinder bore honing.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the cylinder bore honing roughness prediction method according to any one of claims 1 to 8, and the processor is configured to execute the program stored in the memory.

10. A storage medium storing a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the cylinder bore honing roughness prediction method according to any one of claims 1 to 8.

Images