Artificial intelligence-based high-precision control method for bearing grinder

Abstract

The application belongs to the technical field of intelligent control, and particularly relates to a bearing grinder high-precision control method based on artificial intelligence. The method solves the problems of traditional bearing grinder fixed parameter processing, cross-process non-linkage, PID control precision deficiency and the like. The method first constructs and trains three types of models of grinding quality image recognition, grinding parameter sensitivity and cross-process progressive parameter adaptive prediction; through coarse grinding, semi-fine grinding and fine grinding processes, grinding quality characteristics are extracted through the image recognition model and processing errors are generated, error parameter tracing is completed in combination with the sensitivity model, and each process optimal grinding parameter is output by the prediction model; then, the improved grey wolf algorithm of elite level and reverse escape cooperative guidance is used to optimize the PID controller, so as to realize grinding parameter closed-loop control. The application realizes cross-process parameter linkage optimization, improves parameter regulation pertinence and control precision, improves bearing grinding precision consistency, reduces rework rate, and adapts to high-end bearing precision grinding demand.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology, and in particular relates to a high-precision control method for bearing grinding machines based on artificial intelligence. Background Technology

[0002] As a core component of high-end equipment, the grinding precision of bearings directly determines the operational stability and service life of the equipment. With the development of intelligent manufacturing technology, various industries have placed higher demands on the machining precision, process coordination, and parameter adaptive control capabilities of bearing grinding machines. Traditional bearing grinding machine control methods are no longer adequate for the high-precision and high-efficiency grinding requirements, becoming a key bottleneck restricting the manufacturing of high-end bearings. Existing bearing grinding machine grinding control mostly adopts a fixed parameter machining mode, lacking a parameter linkage optimization mechanism between processes; at the same time, PID controller parameter tuning relies heavily on manual experience, and the optimization algorithm is prone to getting trapped in local optima, resulting in insufficient closed-loop control precision. Ultimately, this leads to poor consistency in bearing grinding precision, high process rework rates, and difficulty in meeting the precision grinding standards of high-end bearings. Summary of the Invention

[0003] To address the technical problems existing in the background art described above, this invention proposes a high-precision control method for bearing grinding machines based on artificial intelligence.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] S1. A grinding quality image recognition model, a grinding parameter grinding quality feature sensitivity model, and a cross-process progressive grinding parameter adaptive prediction model are pre-constructed and trained. The grinding quality image recognition model is used to extract grinding quality features from the acquired images of the bearing grinding surface and obtain quality detection results. The grinding parameter grinding quality feature sensitivity model is used to characterize the influence weight of each grinding control parameter on the grinding quality. The cross-process progressive grinding parameter adaptive prediction model is used to predict the optimal grinding control parameters of the subsequent process based on the grinding quality detection results of the previous process.

[0006] S2. The bearing grinding surface is rough ground according to the preset initial grinding control parameters. After the rough grinding is completed, the rough grinding detection image of the bearing grinding surface is collected. The grinding quality features corresponding to the rough grinding process are extracted through the grinding quality image recognition model to generate the rough grinding quality detection result and processing error data.

[0007] S3. Input the rough grinding quality inspection results and machining error data into the cross-process progressive grinding parameter adaptive prediction model. Combine the grinding parameter grinding quality feature sensitivity model to trace the source of rough grinding machining errors and determine the sensitive grinding control parameters corresponding to the errors. At the same time, using the target machining quality of the semi-finish grinding process and the reserved machining allowance threshold of the finish grinding process as constraints, predict the optimal grinding control parameters of the semi-finish grinding process. Complete the semi-finish grinding of the bearing grinding surface according to the optimal semi-finish grinding control parameters. After the semi-finish grinding is completed, collect the semi-finish grinding inspection image of the bearing grinding surface. Extract the grinding quality features corresponding to the semi-finish grinding process through the grinding quality image recognition model to generate the semi-finish grinding quality inspection results and machining error data.

[0008] S4. Input the semi-finish grinding quality inspection results and machining error data into the cross-process progressive grinding parameter adaptive prediction model, and combine it with the grinding parameter grinding quality characteristic sensitivity model to trace the source of semi-finish grinding machining errors and determine the corresponding sensitive grinding control parameters; using the final quality standard of the bearing finished product as a constraint, predict the optimal grinding control parameters for the finishing grinding process; and complete the finishing grinding of the bearing grinding surface according to the optimal finishing grinding control parameters.

[0009] S5. Improved PID control is adopted for the grinding parameters of rough grinding, semi-fine grinding and fine grinding. The closed-loop control corresponds to the actual operating parameters of the actuator, so that the actual grinding parameters in the grinding process track the target grinding control parameters of the process in real time. After the PID closed-loop control of the fine grinding process is completed, the high-precision control of the entire bearing grinding process is completed.

[0010] Preferably, the grinding quality image recognition model is used to extract grinding quality features from the acquired images of the bearing grinding surface and obtain quality inspection results. Specifically, its implementation includes:

[0011] Adaptive preprocessing unit: used to receive the acquired image of the bearing grinding surface, perform noise removal and image grayscale normalization processing, and output a normalized and clear image;

[0012] Multi-scale feature coding unit: used to receive a standardized clear image and obtain a full-dimensional initial feature map covering the global large scale to the micro small scale in the image by improving the dilated convolutional coding branch;

[0013] Standardized feature output unit: used to receive the initial feature map in all dimensions, complete the dimensional mapping and normalization of the features, and output standardized grinding quality features;

[0014] Quality result generation unit: Receives standardized grinding quality features and outputs grinding quality results through a neural network structure.

[0015] Preferably, the implementation of the multi-scale feature coding unit includes:

[0016] The normalized, sharpened image output from the adaptive preprocessing unit is input into the initial convolutional layer to complete the initial mapping from pixel space to high-dimensional feature space, generating the basic feature map corresponding to the single-channel input. Where H and W are the height and width of the feature map, respectively, and C is the initial number of channels; the basic feature map is equally divided into main feature branches along the channel dimension. With auxiliary feature branches ;

[0017] For the main feature branch Four parallel dilated convolutional coding branches are constructed, each corresponding to one of the four scale intervals. The dilated convolutional kernel size of each branch is uniformly 3×3. The dilation rate is adaptively and hierarchically calculated based on the feature scale distribution of the input image. The specific calculation method is as follows: ,in, Let be the adaptive hole rate of the i-th coding branch. Let be the scale coefficient of the i-th branch, with values ​​of 1, 4, 16, and 64, respectively, corresponding to the progression from microscopic small scale to global large scale; The reference pixel size for the bearing grinding surface image; The adaptive calibration coefficients are obtained by mapping the global average pooling output of the auxiliary feature branch through a fully connected layer, with values ​​ranging from [0.8, 1.2]. Each encoding branch takes the main feature branch as input, performs dilated convolution operations with corresponding dilation rates, and outputs four sets of initial scale feature maps with different receptive fields. ;

[0018] Initial scale feature map of the i-th branch First, local feature smoothing is performed using a 3×3 depthwise separable convolution. Then, channel-wise attention modules generate channel-wise weight coefficients, which are then used in conjunction with auxiliary feature branches. The corresponding channel features are weighted and fused to generate a completed continuous-scale feature map. The calculation process is as follows: , , ,in, This is a 3×3 depthwise separable convolution operation. For global average pooling, FC is the fully connected layer mapping operation, and Sigmoid is the activation function. Let be the channel attention weight of the i-th branch, and ⊙ be the Hadamard product operation;

[0019] Finally, the four sets of continuous scale feature maps were completed. The process involves performing adaptive weighted aggregation to generate full-dimensional initial features covering everything from large-scale global features to small-scale micro-features. First, the information entropy of each group of interactive scale feature maps is calculated. Then, the aggregation weights of the corresponding branches are calculated based on the information entropy. The four groups of continuous scale feature maps are weighted and summed based on the aggregation weights. Finally, a 1×1 convolution is performed to unify the channel dimensions, and the full-dimensional initial feature map is output.

[0020] Preferably, the implementation of the standardized feature output unit includes:

[0021] The system receives the initial feature map across all dimensions and performs dual-path spatial feature compression on it. Global average pooling is used to extract the global statistical distribution information of the features, and global max pooling is used to extract the extreme value key information of the features, resulting in a global statistical feature vector. With global extreme value eigenvector After concatenating the two feature vectors along the channel dimension, two fully connected layers are used to complete the high-dimensional feature space mapping and dimension normalization, generating a deep feature vector with fixed dimensions. ,in, These are the mapping operations for the first and second fully connected layers, respectively. For activation function, For channel splicing operations;

[0022] Based on the statistical mean and standard deviation of historical grinding quality characteristics of the current grinding process, the depth feature vector is standardized by Z-score; then, the standardized feature values ​​are constrained to a uniform numerical range of [0,1] by linear mapping to obtain the normalized feature vector, which is used as the final standardized grinding quality feature.

[0023] Preferably, the implementation of the processing error data includes:

[0024] The standardized grinding quality features obtained from the grinding quality image recognition model are decoupled and split according to the feature dimensions of size features, shape and position features, and surface morphology features, respectively, to obtain the measured feature quantization values ​​corresponding to the three types of benchmarks.

[0025] The difference between the measured feature quantification values ​​of each dimension and the preset processing target benchmark value corresponding to the current process is calculated to obtain the processing error data.

[0026] Preferably, the grinding parameter grinding quality characteristic sensitivity model is obtained by constructing a coupled sensitivity matrix, and the construction method includes:

[0027] Through multi-factor, multi-level orthogonal experiments, a dataset of bearing grinding quality characteristics was obtained under different combinations of grinding control parameters. The control parameters include grinding wheel linear speed, grinding feed rate, grinding depth, workpiece spindle speed, and cooling medium flow rate.

[0028] Grey relational analysis was used to calculate the correlation between each grinding control parameter and grinding quality. After normalizing the correlation values, the corresponding sensitivity weights were obtained, and a coupling sensitivity matrix was constructed based on the sensitivity weights.

[0029] Preferably, the cross-process progressive grinding parameter adaptive prediction model is a dual-branch fusion deep learning model, including an error propagation time-series prediction branch and a global parameter optimization reinforcement learning branch.

[0030] The error propagation timing prediction branch is constructed based on a bidirectional long short-term memory network and is used to fit the error propagation law between the three processes of rough grinding, semi-fine grinding and fine grinding. Its input is the grinding quality detection result and processing error data of the previous process, and its output is the error compensation amount required for the subsequent process.

[0031] The global parameter optimization reinforcement learning branch is constructed based on a deep deterministic policy gradient (DDPG) network. It takes the machining allowance threshold and target machining quality of the subsequent process of the bearing grinding machine as hard constraints, and the optimal grinding accuracy of the final product as the reward function. Combined with the error compensation amount and the grinding parameter grinding quality feature sensitivity model, it outputs the optimal grinding control parameters for the subsequent process.

[0032] Preferably, when constructing the continuous motion space in the DDPG network, the sensitive grinding control parameters corresponding to the error are determined by using the grinding parameter grinding quality characteristic sensitivity model. The top three grinding control parameters that contribute the most to the processing error of the previous process are selected, and their unit adjustment step size in the motion space is reduced. For the remaining grinding control parameters, their unit adjustment step size in the motion space is increased.

[0033] Preferably, in step S5, the grinding parameters for rough grinding, semi-fine grinding, and fine grinding are controlled by an improved PID control algorithm using an elite hierarchy and reverse escape collaborative guidance mechanism. Then, a new gray wolf optimization method is used to globally optimize the control parameters of the PID controller to obtain the optimal improved PID controller parameters.

[0034] As a preferred embodiment, the specific implementation of the improved Grey Wolf algorithm using the elite hierarchy and reverse escape collaborative guidance mechanism is as follows:

[0035] Set the gray wolf population size, the maximum number of iterations, and the upper and lower limits of the search space. Randomly initialize the gray wolf population within the search space. The initial convergence factor is 2, and the termination factor is 0.

[0036] Construct a fitness function, calculate the fitness value of each gray wolf in the population, sort all individuals in the population in descending order of fitness value, and obtain the sorted population sequence.

[0037] Calculate the size of the elite pool in the current iteration, and pre-set the maximum size of the elite pool. Where N is the size of the gray wolf population, which decreases progressively with each iteration: , where T is the maximum number of iterations; select the corresponding number of optimal individuals from the sorted population sequence to construct a dynamic elite pool;

[0038] The individual with the best fitness in the dynamic elite pool Based on the upper bound ub and lower bound lb of the search space, the reverse position within the search space is calculated. As a reference position for reverse escape: ;

[0039] The dynamically adjusted backflight weights are calculated based on the current iteration number and the elite pool aggregation degree: ,in, This represents the maximum value of the reverse escape weight. Let be the coefficient of variation of the fitness of all individuals in the dynamic elite pool of generation t, which is obtained by the ratio of the standard deviation of the fitness of all individuals in the dynamic elite pool to the mean fitness.

[0040] The adaptive guidance weight of the dynamic elite pool is used to calculate the positive weighted elite guidance position, and the reverse escape guidance position is calculated by combining the reverse escape reference position. The two guidance positions are weighted and fused by the reverse escape weight to obtain the final gray wolf individual position update formula, and the position update of all individuals in the population is completed.

[0041] After updating the positions of all individuals in the population, update the current iteration count t to t+1; check if the current iteration count has reached the maximum iteration count. If it has, terminate the iteration and output the best individual in the dynamic elite pool as the global optimal solution. If it has not reached the maximum iteration count, return to continue the iteration.

[0042] Compared with existing technologies, the advantages and positive effects of this invention are as follows: It constructs a multi-AI model collaborative system to achieve adaptive prediction and optimization of cross-process parameters, breaking the limitations of traditional fixed-parameter processing and solving the problem of no linkage between processes. The grinding quality image recognition model can accurately extract full-scale grinding features, and combined with a sensitivity model, it enables precise traceability of processing errors, making parameter control more targeted. By improving the Grey Wolf algorithm to optimize PID controller parameters, it avoids the defect of traditional optimization easily getting trapped in local optima, improving closed-loop control accuracy. Simultaneously, full-process detection and real-time parameter tracking significantly improve the consistency of bearing grinding accuracy, reduce process rework rate, and can fully adapt to the precision grinding requirements of high-end bearings. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a structural flowchart of a high-precision control method for bearing grinding machines based on artificial intelligence. Detailed Implementation

[0045] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0046] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0047] This embodiment of the invention provides a high-precision control method for bearing grinding machines based on artificial intelligence, the implementation process of which is as follows: Figure 1 As shown.

[0048] First, a grinding quality image recognition model, a grinding parameter grinding quality feature sensitivity model, and a cross-process progressive grinding parameter adaptive prediction model are pre-constructed and trained. The grinding quality image recognition model is used to extract grinding quality features from the acquired images of the bearing grinding surface and obtain the quality inspection results. The grinding parameter grinding quality feature sensitivity model is used to characterize the influence weight of each grinding control parameter on the grinding quality. The cross-process progressive grinding parameter adaptive prediction model is used to predict the optimal grinding control parameters of the subsequent process based on the grinding quality inspection results of the previous process.

[0049] Furthermore, the grinding quality image recognition model is used to extract grinding quality features from the acquired image of the bearing grinding surface and obtain the quality inspection result. Specifically, its implementation includes:

[0050] The adaptive preprocessing unit receives images of the bearing grinding surface, performs noise removal and image grayscale normalization, and outputs a standardized and clear image. Specifically, it first performs a two-stage cascaded denoising process to address the image interference characteristics of the bearing grinding environment: First, an adaptive median filtering algorithm is used to dynamically adjust the filtering window size, prioritizing the removal of salt-and-pepper noise and impulse interference caused by coolant splashes, grinding wheel debris, and electrical signal interference, while preserving details of the grinding texture and edge contours; then, a bilateral filtering algorithm, combining pixel spatial distance and grayscale similarity as weights, removes Gaussian noise caused by industrial ambient lighting fluctuations, avoiding edge blurring problems caused by traditional Gaussian filtering. After denoising, homomorphic filtering is first used to correct the uneven lighting caused by reflections and shadows on the grinding surface. Then, grayscale normalization is performed to linearly map the image grayscale values ​​to the standard dynamic range of 0 to 255. Z-score normalization is performed based on the grayscale mean and standard deviation of images acquired in the same batch to eliminate grayscale distribution deviations caused by different acquisition conditions and equipment parameters. Finally, a standardized and clear image with uniform pixel size, stable grayscale distribution, and no significant noise is output.

[0051] Multi-scale feature coding unit: Receives a standardized, sharpened image and, by improving the dilated convolutional coding branch, obtains a full-dimensional initial feature map covering the image from global large-scale to micro-scale. Specifically, the input feature initial mapping and dual-channel branch splitting input the standardized, sharpened image output from the adaptive preprocessing unit into the initial convolutional layer, completing the initial mapping from pixel space to high-dimensional feature space and generating the basic feature map corresponding to the single-channel input. Where H and W are the height and width of the feature map, respectively, and C is the initial number of channels; the basic feature map is equally divided into main feature branches along the channel dimension. With auxiliary feature branches The main feature branch is used for multi-scale core feature encoding, while the auxiliary feature branch is used for feature continuity constraints and weight calibration.

[0052] For the main feature branch Four parallel dilated convolutional coding branches are constructed, each representing a different scale range. The dilated convolutional kernel size for each branch is uniformly 3×3. The dilation rate is adaptively and hierarchically calculated based on the feature scale distribution of the input image. The specific calculation method is as follows: ,in, Let be the adaptive hole rate of the i-th coding branch. Let be the scale coefficient of the i-th branch, with values ​​of 1, 4, 16, and 64, respectively, corresponding to the progression from microscopic small scale to global large scale; The reference pixel size for the bearing grinding surface image; The adaptive calibration coefficients are obtained by mapping the global average pooling output of the auxiliary feature branch through a fully connected layer, with values ​​ranging from [0.8, 1.2]. Each encoding branch takes the main feature branch as input, performs dilated convolution operations with corresponding dilation rates, and outputs four sets of initial scale feature maps with different receptive fields. .

[0053] For the initial scale feature map output by each coding branch, a feature continuity completion operation is performed to suppress the inherent grid effect of dilated convolution and complete the semantic relationships between discrete sampling points. Specifically, this is implemented as follows: for the initial scale feature map of the i-th branch... First, local feature smoothing is performed using a 3×3 depthwise separable convolution. Then, channel-wise attention modules generate channel-wise weight coefficients, which are then used in conjunction with auxiliary feature branches. The corresponding channel features are weighted and fused to generate a completed continuous-scale feature map. The calculation process is as follows: , , ,in, This is a 3×3 depthwise separable convolution operation. For global average pooling, FC is the fully connected layer mapping operation, and Sigmoid is the activation function. Let be the channel attention weight for the i-th branch, and ⊙ be the Hadamard product operation. This is applied to the four completed continuous-scale feature maps. The process involves performing adaptive weighted aggregation to generate full-dimensional initial features covering everything from large-scale global features to small-scale micro-features. First, the information entropy of each group of interactive scale feature maps is calculated. Then, the aggregation weights of the corresponding branches are calculated based on the information entropy. The four groups of continuous scale feature maps are weighted and summed based on the aggregation weights. Finally, a 1×1 convolution is performed to unify the channel dimensions, and the full-dimensional initial feature map is output.

[0054] Standardized Feature Output Unit: This unit receives the initial feature map across all dimensions, performs dimensional mapping and normalization of the features, and outputs standardized grinding quality features. Specifically, it receives the initial feature map, performs dual-path spatial feature compression on it, extracts the global statistical distribution information of the features through global average pooling, and extracts the extreme value key information of the features through global max pooling, thus obtaining a global statistical feature vector. With global extreme value eigenvector After concatenating the two feature vectors along the channel dimension, two fully connected layers are used to complete the high-dimensional feature space mapping and dimension normalization, generating a deep feature vector with fixed dimensions. ,in, These are the mapping operations for the first and second fully connected layers, respectively. For activation function, For channel splicing operations; based on the historical grinding quality characteristics statistical mean and standard deviation of the current grinding process, Z-score standardization is performed on the depth feature vector; then, the standardized feature values ​​are constrained to a unified numerical range of [0,1] through linear mapping to obtain the normalized feature vector, which serves as the final standardized grinding quality feature.

[0055] Quality Result Generation Unit: This unit receives standardized grinding quality features and outputs grinding quality results through a neural network structure. Specifically, it constructs a dual-branch fully connected neural network architecture adapted for bearing grinding quality detection. The network body consists of three core fully connected layers. The input dimension of the first layer perfectly matches the dimension of the standardized grinding quality features. The hidden layers use the ReLU activation function to perform nonlinear mapping, and a Dropout layer with a deactivation rate of 0.3 is embedded. The network uses Softmax as the output activation function to output the overall quality score of the bearing grinding surface.

[0056] The grinding parameter grinding quality feature sensitivity model is constructed using a coupled sensitivity matrix. The construction method includes: obtaining a dataset of bearing grinding quality features under different combinations of grinding control parameters through multi-factor, multi-level orthogonal experiments; the control parameters include grinding wheel linear speed, grinding feed rate, grinding depth, workpiece spindle speed, and cooling medium flow rate; using grey relational analysis, calculating the correlation degree between each grinding control parameter and grinding quality, normalizing the correlation degree values ​​to obtain corresponding sensitivity weights, and constructing a coupled sensitivity matrix based on these sensitivity weights. Specifically, the experiment is conducted on a high-precision bearing grinding machine of a specified model, using the inner ring of a deep groove ball bearing as the machining object. Three bearings are repeatedly machined for each parameter combination, and the average value is taken to reduce random errors. After machining, a coordinate measuring machine is used to detect dimensional deviations, roundness, cylindricity, and other form and position features. Quality features are extracted using a grinding quality image recognition model, and all measured quality feature quantification values ​​are compiled into a dataset. Each sample contains the correspondence between control parameters and quality features. In calculating grey relational analysis, a reference sequence and a comparison sequence are constructed. The reference sequence consists of ideal quality feature values, while the comparison sequence consists of measured quality feature values ​​for each experimental group. The data is then dimensionless, and an initialization method is used to divide each sequence by its first value to eliminate the influence of dimensions. Next, the correlation coefficient is calculated, and finally, the average of the correlation coefficients for each parameter is taken to obtain the correlation degree of that parameter on grinding quality. The correlation degrees of the control parameters are summed, and then the correlation degree of each parameter is divided by the sum to obtain the corresponding sensitivity weight. The weight ranges from 0 to 1, and the sum is 1. The larger the weight, the more significant the influence of the parameter on grinding quality. Finally, with grinding control parameters as rows and quality features as columns, the element in the i-th row and j-th column of the matrix represents the sensitivity weight of the i-th control parameter to the j-th type of quality feature, ultimately forming a complete coupled sensitivity matrix for subsequent parameter tracing and cross-process parameter optimization.

[0057] The cross-process progressive grinding parameter adaptive prediction model is a dual-branch fusion deep learning model, including an error propagation time-series prediction branch and a global parameter optimization reinforcement learning branch.

[0058] The error propagation time-series prediction branch is constructed based on a bidirectional long short-term memory network to fit the error propagation law between the three processes of rough grinding, semi-finish grinding, and finish grinding. Its input is the grinding quality inspection results and machining error data of the previous process, and its output is the error compensation amount required for the subsequent process. Specifically, the input data preprocessing and time-series sample construction are completed first: the input data includes the grinding quality inspection results and machining error data of the previous process. First, all input features are normalized to the [0,1] interval by min-max to eliminate dimensional differences; then, the historical machining data are constructed into binary time-series sample pairs according to the process order from rough grinding to semi-finish grinding and from semi-finish grinding to finish grinding. Each sample takes the data of the previous process as input and the theoretical error compensation amount of the subsequent process as label. In terms of network architecture design, a two-layer Bi-LSTM cascaded architecture is adopted, with 64 hidden units in each layer. The bidirectional structure can simultaneously capture the causal relationship of forward error propagation between processes and the inverse correlation of backward quality constraints. A Dropout layer with a deactivation rate of 0.2 is embedded between the two Bi-LSTM layers to suppress overfitting under small sample industrial data. Finally, a fully connected layer is connected to output the error compensation amount corresponding to each quality feature of the subsequent process using a linear activation function. During training, mean squared error is used as the loss function, and the Adam optimizer is used for parameter iteration, ultimately achieving accurate prediction of the subsequent error compensation amount based on the preceding process data.

[0059] Furthermore, the realization of the processing error data involves decoupling and splitting the standardized grinding quality features obtained from the grinding quality image recognition model according to the feature dimensions of size features, shape and position features, and surface morphology features, to obtain the measured feature quantization values ​​corresponding one-to-one with the three types of benchmarks.

[0060] The global parameter optimization reinforcement learning branch is constructed based on a Deep Deterministic Policy Gradient (DDPG) network. It uses the machining allowance threshold and target machining quality of the subsequent process of the bearing grinding machine as hard constraints, and the optimal grinding accuracy of the final product as the reward function. Combining the error compensation amount and the grinding parameter grinding quality feature sensitivity model, it outputs the optimal grinding control parameters for the subsequent process. Furthermore, when constructing the continuous action space in the DDPG network, the sensitive grinding control parameters corresponding to the errors are determined through parameter tracing using the grinding parameter grinding quality feature sensitivity model. The top three grinding control parameters contributing to the machining error of the previous process are selected, and their unit adjustment step size in the action space is reduced. For the remaining grinding control parameters, their unit adjustment step size in the action space is increased. Specifically, the machining error data output from the previous process is organized into an error vector. Using the coupling sensitivity matrix as coefficients, the comprehensive contribution of each parameter is calculated through weighted summation: the contribution of the i-th control parameter is calculated, with a larger contribution indicating a higher weight for the parameter's influence on the current error. The control parameters are sorted in descending order of contribution, and the top three parameters are designated as sensitive control parameters, with the remainder as insensitive control parameters. Then, an initial step size is set based on the grinding machine's adjustment accuracy, and the control parameters are adjusted using preset adjustment coefficients. The adjustment rule is that sensitive control parameters are adjusted to 0.8 times their original step size, and insensitive control parameters are adjusted to 1.2 times their original step size. Using the equipment's allowable upper and lower limits for each parameter as boundaries, and the adjusted step size as the minimum adjustment granularity, a continuous action space for the DDPG is constructed. This ensures both fine-grained control of sensitive parameters and efficient global exploration of insensitive parameters.

[0061] Then, the rough grinding of the bearing surface is completed according to the preset initial rough grinding control parameters. After the rough grinding is completed, the rough grinding detection image of the bearing surface is acquired. The grinding quality features corresponding to the rough grinding process are extracted through the grinding quality image recognition model to generate the rough grinding quality detection results and processing error data.

[0062] The rough grinding quality inspection results and machining error data are input into the cross-process progressive grinding parameter adaptive prediction model. Combined with the grinding parameter grinding quality feature sensitivity model, the parameters of the rough grinding machining error are traced back to determine the sensitive grinding control parameters corresponding to the error. At the same time, the optimal grinding control parameters of the semi-finish grinding process are predicted under the constraints of the target machining quality of the semi-finish grinding process and the reserved machining allowance threshold of the finish grinding process. The semi-finish grinding of the bearing grinding surface is completed according to the optimal semi-finish grinding control parameters. After the semi-finish grinding is completed, the semi-finish grinding inspection image of the bearing grinding surface is collected. The grinding quality features corresponding to the semi-finish grinding process are extracted by the grinding quality image recognition model to generate the semi-finish grinding quality inspection results and machining error data.

[0063] The semi-finish grinding quality inspection results and machining error data are input into the cross-process progressive grinding parameter adaptive prediction model. Combined with the grinding parameter grinding quality characteristic sensitivity model, the semi-finish grinding machining error is traced back to determine the corresponding sensitive grinding control parameters. The optimal grinding control parameters for the fine grinding process are predicted using the final quality standard of the bearing product as a constraint. The fine grinding of the bearing grinding surface is completed according to the optimal fine grinding control parameters.

[0064] An improved PID control is adopted for the grinding parameters of rough grinding, semi-fine grinding and fine grinding. The closed-loop control corresponds to the actual operating parameters of the actuator, so that the actual grinding parameters during the grinding process track the target grinding control parameters of the process in real time. After the PID closed-loop control of the fine grinding process is completed, the high-precision control of the entire bearing grinding process is completed.

[0065] Furthermore, the improved PID control employs an elite hierarchy and reverse escape collaborative guidance mechanism to improve the gray wolf algorithm. Then, a new gray wolf optimization method is used to globally optimize the control parameters of the PID controller to obtain the optimal improved PID controller parameters, thereby achieving precise high-precision control of the bearing grinding machine.

[0066] Furthermore, the specific implementation of the improved gray wolf algorithm using the elite hierarchy and reverse escape collaborative guidance mechanism is as follows: Set the gray wolf population size, the maximum number of iterations, and the upper and lower limits of the search space; randomly initialize the gray wolf population within the search space, with an initial convergence factor of 2 and a termination value of 0; construct a fitness function, calculate the fitness value of each gray wolf individual in the population, and sort all individuals in the population according to their fitness values ​​from largest to smallest to obtain the sorted population sequence; calculate the size of the elite pool for the current iteration, and pre-set the maximum size of the elite pool. Where N is the size of the gray wolf population, which decreases progressively with each iteration: Where T is the maximum number of iterations; select the corresponding number of optimal individuals from the sorted population sequence to construct a dynamic elite pool; select the individual with the best fitness in the dynamic elite pool. Based on the upper bound ub and lower bound lb of the search space, the reverse position within the search space is calculated. As a reference position for reverse escape: The dynamically adjusted backflight weights are calculated based on the current iteration count and the elite pool's aggregation degree. ,in, This represents the maximum value of the reverse escape weight. Let be the coefficient of variation of the fitness of all individuals in the dynamic elite pool in generation t, which is obtained by the ratio of the standard deviation of the fitness of all individuals in the dynamic elite pool to the mean fitness. Based on the adaptive guiding weight of the dynamic elite pool, the forward weighted elite guiding position is calculated, and combined with the reverse escape reference position, the reverse escape guiding position is calculated. The two guiding positions are weighted and fused using the reverse escape weight to obtain the final gray wolf individual position update formula, completing the position update of all individuals in the population. After completing the position update of all individuals in the population, the current iteration number t is updated to t+1. It is determined whether the current iteration number has reached the maximum iteration number. If it has, the iteration is terminated, and the best individual in the dynamic elite pool is output as the global optimal solution; otherwise, the iteration continues.

[0067] Furthermore, the positively weighted elite guidance position is calculated based on the adaptive guidance weight of the dynamic elite pool, and the reverse escape guidance position is calculated by combining the reverse escape baseline position. The two guidance positions are then weighted and fused using the reverse escape weight to obtain the final gray wolf individual position update formula. The specific implementation of updating the position of all individuals in the population is as follows: First, the convergence factor A and the random perturbation factor C are calculated: ,in, These are the initial and final values ​​of the convergence factor, respectively. , ,in, Let be the convergence factor of the current iteration. The number is a uniformly random number between 0 and 1; the guiding candidate position is calculated for each individual in the dynamic elite pool, and for the i-th individual in the elite pool... The candidate positions it guides Calculation formula: ,in, Given the position of the gray wolf individual in the current iteration; calculate the positively weighted elite guidance position. The candidate positions of all elite individuals are summed using a weighted average based on the guiding weight: The guiding weight is obtained based on the fitness ranking of individuals within the dynamic elite pool; then the reverse escape guiding position is calculated. Finally, the final location update of the individual gray wolf is calculated to complete the location update. The calculation formula is as follows: ,in, Position of the gray wolf individual for the next iteration.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A high-precision control method for bearing grinding machines based on artificial intelligence, characterized in that, Includes the following steps: S1. Pre-construct and train the grinding quality image recognition model, the grinding parameter grinding quality feature sensitivity model, and the cross-process progressive grinding parameter adaptive prediction model. S2. Complete the rough grinding of the bearing grinding surface according to the preset initial parameters. After rough grinding, collect the grinding surface detection image. Extract the rough grinding quality features through the grinding quality image recognition model and generate the corresponding rough grinding quality detection results and processing error data. S3. Input the rough grinding quality inspection results and processing error data into the adaptive prediction model, combine the sensitivity model to trace the source of the rough grinding error parameters and determine the sensitive parameters; use the semi-finish grinding target processing quality and the fine grinding reserved processing allowance threshold as constraints to predict the optimal semi-finish grinding parameters and complete the processing; then collect images and generate semi-finish grinding quality inspection results and processing error data through the recognition model. S4. Input the semi-finish grinding quality inspection results and machining error data into the adaptive prediction model, combine the sensitivity model to trace the source of semi-finish grinding error parameters and determine the sensitive parameters; use the final quality standard of the bearing finished product as a constraint to predict the optimal parameters for fine grinding and complete the fine grinding of the bearing grinding surface; S5. Improved PID control is adopted for the grinding parameters of rough grinding, semi-fine grinding and fine grinding. The closed-loop control corresponds to the actual operating parameters of the actuator, so that the actual grinding parameters in the grinding process track the target grinding control parameters of the process in real time. After the PID closed-loop control of the fine grinding process is completed, the high-precision control of the entire bearing grinding process is completed.

2. The high-precision control method for bearing grinding machines based on artificial intelligence according to claim 1, characterized in that, The specific implementation of the grinding quality image recognition model includes: Adaptive preprocessing unit: used to receive the acquired image of the bearing grinding surface, perform noise removal and image grayscale normalization processing, and output a normalized and clear image; Multi-scale feature coding unit: used to receive a standardized clear image and obtain a full-dimensional initial feature map covering the global large scale to the micro small scale in the image by improving the dilated convolutional coding branch; Standardized feature output unit: used to receive the initial feature map in all dimensions, complete the dimensional mapping and normalization of the features, and output standardized grinding quality features; Quality result generation unit: Receives standardized grinding quality features and outputs grinding quality results through a neural network structure.

3. The high-precision control method for bearing grinding machines based on artificial intelligence according to claim 2, characterized in that, The implementation of the multi-scale feature coding unit includes: The normalized, sharpened image output from the adaptive preprocessing unit is input into the initial convolutional layer to complete the initial mapping from pixel space to high-dimensional feature space, generating the basic feature map corresponding to the single-channel input. Where H and W are the height and width of the feature map, respectively, and C is the initial number of channels; the basic feature map is equally divided into main feature branches along the channel dimension. With auxiliary feature branches ; For the main feature branch Four parallel dilated convolutional coding branches are constructed, each corresponding to one of the four scale intervals. The dilated convolutional kernel size of each branch is uniformly 3×3. The dilation rate is adaptively and hierarchically calculated based on the feature scale distribution of the input image. The specific calculation method is as follows: ,in, Let be the adaptive hole rate of the i-th coding branch. Let be the scale coefficient of the i-th branch, with values ​​of 1, 4, 16, and 64, respectively, corresponding to the progression from microscopic small scale to global large scale; The reference pixel size for the bearing grinding surface image; The adaptive calibration coefficients are obtained by mapping the global average pooling output of the auxiliary feature branch through a fully connected layer, with values ​​ranging from [0.8, 1.2]. Each encoding branch takes the main feature branch as input, performs dilated convolution operations with corresponding dilation rates, and outputs four sets of initial scale feature maps with different receptive fields. ; Initial scale feature map of the i-th branch First, local feature smoothing is performed using a 3×3 depthwise separable convolution. Then, channel-wise attention modules generate channel-wise weight coefficients, which are then used in conjunction with auxiliary feature branches. The corresponding channel features are weighted and fused to generate a completed continuous-scale feature map. The calculation process is as follows: , , ,in, This is a 3×3 depthwise separable convolution operation. For global average pooling, FC is the fully connected layer mapping operation, and Sigmoid is the activation function. Let be the channel attention weight of the i-th branch, and ⊙ be the Hadamard product operation; Finally, the four sets of continuous scale feature maps were completed. The process involves performing adaptive weighted aggregation to generate full-dimensional initial features covering everything from large-scale global features to small-scale micro-features. First, the information entropy of each group of interactive scale feature maps is calculated. Then, the aggregation weights of the corresponding branches are calculated based on the information entropy. The four groups of continuous scale feature maps are weighted and summed based on the aggregation weights. Finally, a 1×1 convolution is performed to unify the channel dimensions, and the full-dimensional initial feature map is output.

4. The high-precision control method for bearing grinding machines based on artificial intelligence according to claim 2, characterized in that, The implementation of the standardized feature output unit includes: The system receives the initial feature map across all dimensions and performs dual-path spatial feature compression on it. Global average pooling is used to extract the global statistical distribution information of the features, and global max pooling is used to extract the extreme value key information of the features, resulting in a global statistical feature vector. With global extreme value eigenvector After concatenating the two feature vectors along the channel dimension, two fully connected layers are used to complete the high-dimensional feature space mapping and dimension normalization, generating a deep feature vector with fixed dimensions. ,in, These are the mapping operations for the first and second fully connected layers, respectively. For activation function, For channel splicing operations; Based on the statistical mean and standard deviation of historical grinding quality characteristics of the current grinding process, the depth feature vector is standardized by Z-score; then, the standardized feature values ​​are constrained to a uniform numerical range of [0,1] by linear mapping to obtain the normalized feature vector, which is used as the final standardized grinding quality feature.

5. The high-precision control method for bearing grinding machines based on artificial intelligence according to claim 1, characterized in that, The implementation of the processing error data includes: The standardized grinding quality features obtained from the grinding quality image recognition model are decoupled and split according to the feature dimensions of size features, shape and position features, and surface morphology features, respectively, to obtain the measured feature quantization values ​​corresponding to the three types of benchmarks. The difference between the measured feature quantification values ​​of each dimension and the preset processing target benchmark value corresponding to the current process is calculated to obtain the processing error data.

6. The high-precision control method for bearing grinding machines based on artificial intelligence according to claim 1, characterized in that, The grinding parameter grinding quality characteristic sensitivity model is obtained by constructing a coupled sensitivity matrix, and its construction method includes: Through multi-factor, multi-level orthogonal experiments, a dataset of bearing grinding quality characteristics was obtained under different combinations of grinding control parameters. The control parameters include grinding wheel linear speed, grinding feed rate, grinding depth, workpiece spindle speed, and cooling medium flow rate. Grey relational analysis was used to calculate the correlation between each grinding control parameter and grinding quality. After normalizing the correlation values, the corresponding sensitivity weights were obtained, and a coupling sensitivity matrix was constructed based on the sensitivity weights.

7. The high-precision control method for bearing grinding machines based on artificial intelligence according to claim 1, characterized in that, The cross-process progressive grinding parameter adaptive prediction model is a deep learning model with dual branches, including an error propagation time-series prediction branch and a global parameter optimization reinforcement learning branch. The error propagation timing prediction branch is constructed based on a bidirectional long short-term memory network and is used to fit the error propagation law between the three processes of rough grinding, semi-fine grinding and fine grinding. Its input is the grinding quality detection result and processing error data of the previous process, and its output is the error compensation amount required for the subsequent process. The global parameter optimization reinforcement learning branch is constructed based on a deep deterministic policy gradient (DDPG) network. It takes the machining allowance threshold and target machining quality of the subsequent process of the bearing grinding machine as hard constraints, and the optimal grinding accuracy of the final product as the reward function. Combined with the error compensation amount and the grinding parameter grinding quality feature sensitivity model, it outputs the optimal grinding control parameters for the subsequent process.

8. The high-precision control method for bearing grinding machines based on artificial intelligence according to claim 7, characterized in that, When constructing a continuous motion space in the DDPG network, the sensitive grinding control parameters corresponding to the error are determined by using the grinding parameter grinding quality characteristic sensitivity model. The top three grinding control parameters that contribute the most to the processing error of the previous process are selected, and their unit adjustment step size in motion space is reduced. For the remaining grinding control parameters, their unit adjustment step size in motion space is increased.

9. The high-precision control method for bearing grinding machines based on artificial intelligence according to claim 1, characterized in that, In step S5, the grinding parameters for rough grinding, semi-fine grinding, and fine grinding are controlled by an improved PID control. This involves using an elite hierarchy and reverse escape collaborative guidance mechanism to improve the gray wolf algorithm. Then, a new gray wolf optimization method is used to globally optimize the control parameters of the PID controller to obtain the optimal improved PID controller parameters.

10. A high-precision control method for a bearing grinding machine based on artificial intelligence according to claim 9, characterized in that, The specific implementation of the improved Grey Wolf algorithm using the elite hierarchy and reverse escape collaborative guidance mechanism is as follows: Set the gray wolf population size, the maximum number of iterations, and the upper and lower limits of the search space. Randomly initialize the gray wolf population within the search space. The initial convergence factor is 2, and the termination factor is 0. Construct a fitness function, calculate the fitness value of each gray wolf in the population, sort all individuals in the population in descending order of fitness value, and obtain the sorted population sequence. Calculate the size of the elite pool in the current iteration, and pre-set the maximum size of the elite pool. Where N is the size of the gray wolf population, which decreases progressively with each iteration: , where T is the maximum number of iterations; select the corresponding number of optimal individuals from the sorted population sequence to construct a dynamic elite pool; The individual with the best fitness in the dynamic elite pool Based on the upper bound ub and lower bound lb of the search space, the reverse position within the search space is calculated. As a reference position for reverse escape: ; The dynamically adjusted backflight weights are calculated based on the current iteration number and the elite pool aggregation degree: ,in, This represents the maximum value of the reverse escape weight. Let be the coefficient of variation of the fitness of all individuals in the dynamic elite pool of generation t, which is obtained by the ratio of the standard deviation of the fitness of all individuals in the dynamic elite pool to the mean fitness. The adaptive guidance weight of the dynamic elite pool is used to calculate the positive weighted elite guidance position, and the reverse escape guidance position is calculated by combining the reverse escape reference position. The two guidance positions are weighted and fused by the reverse escape weight to obtain the final gray wolf individual position update formula, and the position update of all individuals in the population is completed. After updating the positions of all individuals in the population, update the current iteration count t to t+1; check if the current iteration count has reached the maximum iteration count. If it has, terminate the iteration and output the best individual in the dynamic elite pool as the global optimal solution. If it has not reached the maximum iteration count, return to continue the iteration.

Images