Metal processing process parameter intelligent recommendation method and system based on big data

Abstract

This invention relates to the field of machine learning technology, solving the technical problems of existing technologies being prone to getting trapped in local optima, having high computational costs, being difficult to accurately construct models, and being unable to achieve safe and efficient parameter optimization under limited resources. In particular, it relates to an intelligent recommendation method and system for metal processing process parameters based on big data. The method involves the following steps: collecting the original process data stream of the metal processing system; generating integrated process data by integrating and processing the original process data stream; constructing reinforcement learning states based on clustering context vectors, which enhances the baseline reliability and security of the recommendations; performing probabilistic modeling and uncertainty quantification for unknown regions; and demonstrating excellent adaptability and robustness through a weighted fusion formula for multi-objective optimization fusion, avoiding catastrophic failures while actively seeking optimal solutions, thereby generating optimal process parameter recommendations efficiently and reliably.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a metal processing process parameter intelligent recommendation method and system based on big data. BACKGROUND

[0002] Machine learning is a core branch of artificial intelligence, which belongs to a subclass of big data analysis, and it studies how to let computer systems automatically improve algorithm performance using data. The core is to analyze expressions from data and use the expressions to predict unknown situations. In the metal processing process parameter intelligent recommendation method based on machine learning, the prior art usually manually selects and constructs important feature variables from the collected original process data, selects a traditional machine learning algorithm according to the problem type, inputs the processed feature data and historical processing results into the algorithm as training samples, learns the mapping relationship between the process parameters and the processing results through iterative training, generates a prediction model, and finally uses the model to predict the parameters of new process requirements and complete the recommendation.

[0003] The prior art simply uses reinforcement learning based on historical experience. When facing a new process scene with sparse data, it is easy to fall into local optimization. When facing a high-dimensional and complex process parameter space, the calculation cost is high, and the model is difficult to accurately construct. In practice, it is easy to encounter special high-risk scenes such as first-piece processing with special materials or process limit breakthrough. At this time, it is often necessary to find feasible or even optimal parameters under the condition of very few trial and error. The prior art either recommends mediocre solutions due to excessive conservatism or leads to uncontrollable trial and error costs due to blind exploration, making it difficult to achieve safe and efficient parameter optimization under limited resources. SUMMARY

[0004] In view of the shortcomings of the prior art, the present application provides a metal processing process parameter intelligent recommendation method and system based on big data, which solves the technical problems that the prior art is easy to fall into local optimization, the calculation cost is high, the model is difficult to accurately construct, and it is difficult to achieve safe and efficient parameter optimization under limited resources.

[0005] To solve the above technical problems, the present application provides the following technical scheme: a metal processing process parameter intelligent recommendation method based on big data, the steps of the method are: collecting original process data stream of a metal processing system, generating integrated process data according to the original process data stream through integration processing; Performing covariance decomposition processing based on the integrated process data to obtain an eigenvalue vector and an eigenvector matrix, performing dimensionality reduction coding processing on the eigenvalue vector and the eigenvector matrix to obtain feature process data; Performing density clustering on the feature process data to obtain a clustering label vector, performing isolated forest anomaly processing and anomaly clustering integration based on the clustering label vector to generate clustered process data; According to the clustering process data, a reinforcement learning action is obtained through strategy inference processing, multi-objective optimization is performed based on the clustering process data and the reinforcement learning action, and a recommended process parameter set is generated; The recommended process parameter set is verified and fine-tuned to generate optimized process parameter data.

[0006] Preferably, according to the original process data stream, an integration processing includes: performing multi-source data processing based on the original process data stream to generate an original process data set; According to the original process data set, an outlier and missing value processing is performed to obtain a cleaning process data set; The cleaning process data set is standardized and merged to obtain an integrated process data.

[0007] Preferably, a covariance decomposition processing is performed based on the integrated process data, including: performing a standardization processing based on the integrated process data to obtain a standardized process data set; According to the standardized process data set, a covariance processing is performed to generate a covariance matrix; The covariance matrix is processed by eigenvalue decomposition to generate an eigenvalue vector and an eigenvector matrix.

[0008] Preferably, the eigenvalue vector and the eigenvector matrix are processed by dimensionality reduction encoding, including: performing dimensionality reduction projection on the standardized process data set, the eigenvalue vector and the eigenvector matrix to obtain a dimensionality reduction process data; Based on the dimensionality reduction process data, an encoding feature processing is performed to generate an encoding feature vector; According to the encoding feature vector, a feature selection processing is performed to obtain a feature process data.

[0009] Preferably, the feature process data is processed by density clustering, including: according to the feature process data, a parameter estimation processing is performed to obtain a field parameter and a minimum point number; Based on the feature process data, the field parameter and the minimum point number, a core point detection is performed to generate a core point set; The core point set, the feature process data and the field parameter are processed by density clustering to obtain a clustering label vector.

[0010] Preferably, based on the clustering label vector, an isolated forest anomaly processing and an anomaly clustering integration are performed, including: based on the feature process data, an isolated forest processing is performed to obtain an isolated tree set; The isolated tree set and the feature process data are processed by path length to obtain an anomaly score vector; The anomaly score vector and the clustering label vector are integrated by anomaly clustering to generate a clustering process data.

[0011] Preferably, the clustering process data is processed through strategy reasoning, including: extracting contextual features based on the clustering process data to obtain a clustering context vector; Reinforcement learning states are generated by processing the clustering context vectors through state encoding. The reinforcement learning state is processed by policy reasoning to obtain the reinforcement learning action.

[0012] Preferably, multi-objective optimization is performed based on clustering process data and reinforcement learning actions, including: extracting historical process parameters and performance data from historical clustering process data, performing Gaussian regression processing on the historical process parameters and performance data, and obtaining a Gaussian processing model; Based on the Gaussian processing model and reinforcement learning state, the desired improvement processing is performed to obtain the acquisition function value; Based on the collected function values ​​and reinforcement learning actions, a set of recommended process parameters is generated through multi-objective optimization.

[0013] Preferably, the recommended process parameter set is verified and fine-tuned, including: constructing a digital twin based on the recommended process parameter set to generate simulated performance data; The evaluation results are obtained by processing the simulated performance data through key indicators; The evaluation results are fine-tuned to obtain the optimal process parameter data.

[0014] The technical solution also provides a big data-based intelligent recommendation system for metal processing parameters. This system includes an integration module, which collects the raw process data stream of the metal processing system and generates integrated process data through integration processing based on the raw process data stream. The feature module is used to perform covariance decomposition based on integrated process data to obtain eigenvalue vectors and eigenvector matrices, and then perform dimensionality reduction encoding on the eigenvalue vectors and eigenvector matrices to obtain feature process data. The clustering module is used to perform density clustering on the feature process data to obtain cluster label vectors. Based on the cluster label vectors, isolated forest anomaly processing and anomaly cluster integration are performed to generate clustered process data. The recommended parameter module is used to obtain reinforcement learning actions by processing the clustered process data through policy reasoning, and to perform multi-objective optimization based on the clustered process data and reinforcement learning actions to generate a recommended process parameter set. The optimal parameter module is used to verify and fine-tune the recommended process parameter set to generate optimal process parameter data.

[0015] By employing the above technical solution, the present invention provides a method and system for intelligent recommendation of metal processing parameters based on big data, which has at least the following beneficial effects: 1. This invention performs statistical preprocessing on the integrated process data through standardization and principal component analysis, which can efficiently remove redundant information and random noise, significantly improving the training stability of subsequent models. Through the autoencoder, there is no need to predefine the mechanism model. It learns autonomously and extracts potential effective representations in a data-driven manner. The feature process data has high information density and strong discriminative power, effectively solving the black box problem of process mechanism and laying a high-quality data foundation for subsequent intelligent recommendation models.

[0016] 2. This invention reduces subjective dependence by using automated methods such as distance maps, enabling it to adaptively discover density clusters of arbitrary shapes and sizes in the feature space. The isolated forest uses a random partitioning strategy to evaluate the segregability of each sample point from another dimension. Based on the initial clustering result labels, it uses anomaly scores to perform fine-tuning, achieving a balance between noise filtering and edge information preservation, significantly improving the purity and completeness of the clustering results. When facing unconventional scenarios such as special materials, this combined method exhibits strong robustness, resulting in higher quality clustering process data and providing high-quality and stable data for subsequent steps.

[0017] 3. This invention constructs reinforcement learning states based on clustering context vectors, enhancing the baseline reliability and security of recommendations. Based on a Bayesian optimization module using Gaussian processes, it performs probability modeling and uncertainty quantification for unknown regions. Multi-objective optimization fusion uses a weighted fusion formula to comprehensively score each recommendation. When facing black-box scenarios such as first-piece processing of special materials or breakthroughs in process limits, this combined method demonstrates excellent adaptability and robustness, avoiding catastrophic failures while actively seeking optimal solutions, thus generating optimal process parameter recommendations efficiently and reliably.

[0018] 4. This invention achieves improved efficiency and accuracy through virtual verification and optimization using digital twin simulation, comprehensive evaluation, and gradient feedback fine-tuning. It utilizes a high-fidelity physical model to safely, quickly, and cost-effectively predict processing results in digital space, avoiding material waste, equipment damage, and production delay risks. Through comprehensive evaluation of multiple indicators with configurable weights, it flexibly adapts to different production optimization goals. The introduction of a gradient-based automatic feedback adjustment mechanism enables automatic and accurate parameter correction based on simulation performance, significantly improving the reliability and optimality of the final recommended parameters. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the intelligent recommendation method for metal processing parameters based on big data according to the present invention. Figure 2 This is a structural block diagram of the intelligent recommendation system for metal processing parameters based on big data, as described in this invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0021] Example 1: Due to the inherent limitations of existing technologies—their susceptibility to local optima, high computational costs, and difficulty in accurately constructing models—it remains challenging to achieve safe and efficient parameter optimization within limited resources. Please refer to [the relevant documentation / reference]. Figure 1 This embodiment provides an intelligent recommendation method for metal processing parameters based on big data, which can avoid getting trapped in local optima, save costs, and achieve efficient and safe parameter optimization. The method includes the following steps: S1. Collect the raw process data stream of the metal processing system, and generate integrated process data through integration processing based on the raw process data stream; When processing similar metal processing process data, existing technologies often rely on a single or a few data sources for data collection, which cannot fully reflect the overall process. Data cleaning often adopts methods based on fixed rules or human experience judgment. This approach is prone to accidentally deleting valid data or retaining outliers, and it cannot adapt to changes in different parameter characteristics. The processing results are unstable and have poor repeatability. Data from different sources are only physically stacked, lacking unified format specifications and dimensional standardization processing. To solve the above problems, the specific implementation steps are as follows: S11. Perform multi-source data processing based on the original process data stream to generate the original process dataset. In this step, the input is the original process data stream from multiple channels, such as processing temperature, pressure and speed data monitored by real-time sensors, successful or failed process parameter records accumulated in the historical database, and metal hardness, toughness and other attribute information provided in the material library. The processing adopts multi-source fusion acquisition, and the steps are as follows: First, all data points from the first data source are used as the basic set. Then, data points from the second data source are added to the basic set one by one, which is equivalent to superimposing and merging on the original data. Then, data points from the third data source are introduced and further combined with the integrated dataset. And so on, gradually incorporating data points from all data sources. For example, if there are multiple data sources, the original process dataset is equal to the data from the first data source plus the data from the second data source plus the data from the third data source, until the last data source. Finally, through this successive merging method, a complete data set covering all source information is formed.

[0022] S12. Based on the original process dataset, outliers and missing values ​​are processed to obtain the cleaning process dataset. In this step, for each feature in the original process dataset, such as cutting depth or heat treatment time, the mean is calculated. The mean is calculated by adding all valid data points of the feature one by one, and then dividing by the total number of data points. For example, if there are ten data points, the sum is divided by ten to get the mean. Next, the standard deviation of the feature is calculated to measure the degree of data dispersion. The calculation is to first find the difference between each data point and the mean, square these differences respectively, then add all the squared values, divide by the total number of data points, and finally take the square root of the result to get the standard deviation. Outlier detection is performed based on the mean and standard deviation: For each data point, the absolute value of its difference from the mean is calculated. If this absolute value is greater than three times the standard deviation, it is considered an outlier. For example, if the mean is 100, the standard deviation is 5, and three times the standard deviation is 15, then if the difference between data point 120 and the mean is 20, which is greater than 15, it is considered an outlier. Outliers are either removed or replaced with the median of the feature. The median is obtained by taking the middle value after sorting all data points. At the same time, for missing values, i.e., data points that are empty, they are directly filled with the mean of the feature. Through the above calculations, each data point is processed: missing values ​​are filled with the mean, outliers are corrected with the median, and normal values ​​are retained. Finally, the cleaning process dataset is obtained, which removes noise and errors from the original data and ensures the consistency and authenticity of the data.

[0023] S13. Standardize and merge the cleaning process dataset to obtain integrated process data. In this step, the cleaning process dataset has removed outliers and missing values. It includes real-time processing temperature, pressure, and speed monitored by sensors, as well as process parameters and material properties from historical records. First, based on the key values ​​of the process parameters, for example, using the names of parameters such as cutting depth, spindle speed, and heat treatment temperature as common identifiers, the multi-source data from sensor streams, databases, etc., are associated and merged according to the same key value to ensure that all values ​​of each parameter are integrated into the same framework, eliminating duplicate or conflicting entries, and forming a structurally consistent intermediate dataset. Then, data standardization is performed to eliminate the differences in dimensions and scales between different parameters. For each process parameter, such as milling speed, all original values ​​of the parameter are extracted from the cleaning dataset. First, the minimum and maximum values ​​among these values ​​are found. Then, normalization calculation is performed on each original value: First, the minimum value is subtracted from the original value to obtain the deviation between the original value and the minimum value. Second, the difference between the maximum value and the minimum value is calculated, which is the numerical range of the parameter. Third, the deviation is divided by the range to obtain the standardized value, which is the integrated process data.

[0024] This invention overcomes the limitations of traditional methods, such as single data sources and incomplete information dimensions, by using multi-source fusion acquisition. It constructs a comprehensive and three-dimensional original data foundation for subsequent analysis. The data cleaning step adopts automated anomaly detection and intelligent filling based on statistical principles, which greatly improves the objectivity and reliability of data cleaning, effectively preserves valuable data features, and removes noise interference. The data integration step maps different process parameters to the same standard scale through key-value merging and maximum-minimum normalization.

[0025] S2. Based on the integrated process data, covariance decomposition is performed to obtain eigenvalue vectors and eigenvector matrices. The eigenvalue vectors and eigenvector matrices are then subjected to dimensionality reduction encoding to obtain characteristic process data. Existing technologies are susceptible to multicollinearity and random noise interference, resulting in unstable training processes, difficulty in convergence, and a tendency to overfit training samples, leading to poor generalization ability. Whether it is a fixed linear transformation trained based on historical data or a handcrafted feature designed based on old mechanisms, it will completely fail due to mode mismatch. For example, when facing micro-nano scale cutting or ultra-high temperature instantaneous forming processes, the physical processes are complex and lack clear theoretical models. There are no reference processing methods in historical data, resulting in the black box problem of process mechanisms. To solve the above problems, the specific implementation steps are as follows: S21. Standardize the integrated process data to obtain a standardized process dataset. This step integrates process data including key process parameters such as turning speed, feed rate, depth of cut, heat treatment temperature, and cooling time. First, for a selected feature, such as turning speed, calculate the arithmetic mean of all its data points. The calculation method is to add each value under this feature one by one to obtain the sum, and then divide this sum by the total number of data points. Next, calculate the standard deviation of this feature to measure the dispersion of the data. During the calculation, first find the difference between each data point and the calculated mean, and then... Square each difference individually, then sum all the squared values, divide this sum by the total number of data points, and finally take the square root of the quotient. After calculating the mean and standard deviation, standardize each original data point using the following formula: subtract the mean of the feature from the original value to obtain a centered difference, then divide this difference by the standard deviation of the feature. The resulting standardized process dataset transforms parameters with different dimensions and distribution ranges, such as pressure and speed, to the same comparable statistical scale, which is beneficial for subsequent steps.

[0026] S22. Perform covariance processing on the standardized process dataset to generate a covariance matrix. In this step, the standardized process dataset is first regarded as a mathematical data matrix, where each row of the matrix represents a processing sample and each column represents a specific process parameter feature. Then, in order to calculate the covariance matrix, matrix operations need to be performed: The first step is to transpose the data matrix, that is, the rows of the original matrix become the columns of the new matrix, and the columns of the original matrix become the rows of the new matrix, thus obtaining a transposed matrix. The second step is to multiply this transposed matrix with the original data matrix. The multiplication rule is to take each row of the transposed matrix corresponding to each column feature of the original data matrix and perform a dot product calculation with each column feature of the original data matrix. That is, multiply the two values ​​at corresponding positions and add all the product results to obtain a new value. This value reflects the preliminary value of the covariance between the two features. The third step involves dividing each value in the new matrix obtained in the previous step by the total number of samples to obtain the final covariance matrix. This covariance matrix is ​​a symmetric matrix, with its rows and columns corresponding to various process parameter features. The elements on the diagonal of the matrix are the variance of each feature itself, while the elements off-diagonal represent the covariance between different features pairwise. Positive values ​​indicate that two features tend to increase or decrease together, negative values ​​indicate that when one feature increases, the other decreases, and zero values ​​indicate no linear correlation. The covariance matrix is ​​the core foundation for subsequent principal component analysis. By performing feature decomposition on this matrix, the most important linear combinations in the data, namely the principal components, can be extracted, thereby achieving dimensionality reduction in intelligent recommendation of metal processing parameters.

[0027] S23. Perform eigenvalue decomposition on the covariance matrix to generate eigenvalue vectors and eigenvector matrices. This step first establishes a characteristic equation, which is expressed as the covariance matrix multiplied by an eigenvector to be determined, equal to a specific eigenvalue scalar multiplied by the same eigenvector. Solving this equation requires a series of mathematical operations. Specifically, the first step is to construct a characteristic polynomial related to the covariance matrix. This is done by multiplying an unknown by the identity matrix, subtracting the result from the covariance matrix, calculating the determinant of the new matrix, and setting the determinant to zero, forming a higher-order equation about the unknown. The second step is to solve this characteristic polynomial equation, i.e., to find all roots that make the determinant zero. These roots are the eigenvalues ​​we are looking for. Each eigenvalue is a real number, and their magnitude represents the variance of the data in the corresponding new direction, i.e., the importance of that direction. The third step is to substitute each obtained eigenvalue back into the original characteristic equation. This requires subtracting the product of the eigenvalue and the identity matrix from the covariance matrix, and then solving the homogeneous linear equation system formed to find a non-zero vector solution. This vector solution is the eigenvector corresponding to the eigenvalue, which defines a new coordinate axis direction in the data space. Finally, all eigenvalues ​​are arranged in descending order of value to form the eigenvalue vector, and all corresponding eigenvectors are arranged as column vectors in the order of their eigenvalues ​​to form the eigenvector matrix.

[0028] S24. Perform dimensionality reduction projection on the standardized process dataset, eigenvalue vectors, and eigenvector matrix to obtain dimensionality-reduced process data. In this step, principal component selection is performed first. Based on the eigenvalue vectors, the top k largest eigenvalues ​​are selected. The specific number of k is usually determined based on the cumulative variance contribution rate. The k eigenvectors corresponding to these k largest eigenvalues ​​are extracted from the eigenvector matrix and combined column by column to form a new projection matrix. Each column of this matrix represents a selected principal component direction. Then, projection dimensionality reduction calculation is performed. The standardized process data is regarded as a data matrix, where each row is a processing sample and each column is a standardized process parameter. The projection operation is to multiply this data matrix with the newly created projection matrix. The specific process of multiplication is as follows: for each sample row vector in the data matrix, a dot product operation is performed with each principal component eigenvector in the projection matrix. During the dot product calculation, each value of the sample vector is multiplied by the value at the corresponding position of the eigenvector, and then all these product results are added together to obtain a scalar value. This value is the coordinate score of the sample in the current principal component direction. This operation traverses all samples and all selected k principal components, and finally generates a new data matrix, namely the dimensionality-reduced process dataset. The number of samples in this dataset is the same as the original data, but each sample is now composed of only k principal component scores. The number of features is greatly reduced from the original dozens or hundreds of process parameters. These principal component features are orthogonal to each other, uncorrelated, and ordered by importance. In the intelligent recommendation of metal processing parameters, the dimensionality-reduced process dataset compresses the complex multidimensional parameter space, such as rotational speed, feed, temperature, and pressure, into a few comprehensive indicators, effectively eliminating redundant collinearity between the original parameters and filtering out secondary noise.

[0029] S25. Based on the dimensionality reduction process data, perform encoding feature processing to generate encoded feature vectors. In this step, for each sample in the input data, i.e., each vector composed of several principal component scores, the encoder performs a forward propagation calculation. The first step of the forward propagation calculation is a linear transformation: the encoding neural network layer contains a set of adjustable weight parameters and a bias parameter. The linear transformation operation is to multiply each value of the input sample vector by a set of weight parameters corresponding to a certain encoding neuron in that layer, then sum all these products, and finally add the bias parameter value corresponding to that neuron. This calculation process is performed independently for each neuron in that encoding layer. First, a preliminary linear weighted sum is generated for each neuron. The second step is nonlinear activation: the linear weighted sum calculated for each neuron in the previous step is input into an activation function called a linear rectified function. The function is defined as follows: if the input value is greater than zero, the value is output as is; if the input value is less than or equal to zero, the output is zero. Through this nonlinear operation, the network can learn and express complex nonlinear relationships in the data. The set of values ​​obtained after this layer of calculation constitutes the feature representation of the sample in the encoder output layer, that is, an encoded feature vector. This process traverses all input samples and finally obtains the set of encoded feature vectors corresponding to all samples.

[0030] S26. Based on the encoded feature vector, feature selection processing is performed to obtain feature process data; the input in this step is the set of encoded feature vectors calculated by the encoder forward propagation in the previous step. Each vector in this vector set is a deep abstract representation of the original process data after a series of complex nonlinear transformations. The specific processing is direct transmission. Since the previous encoding step has already completed the deep compression and refinement of information through a neural network, its output is an encoded feature vector, which already has the characteristics of high information density and strong representation ability. Therefore, in this step, there is no need to perform calculations on these vectors, nor is there any need for secondary selection or combination. The entire set of input encoded feature vectors is specified as the final output data without modification. The feature process data is highly referential data after the original multi-source process parameters have been refined by both linear and nonlinear methods. In the intelligent recommendation of metal processing process parameters, this data is no longer a simple raw reading such as rotation speed and temperature, nor is it a linear principal component score. Instead, it is a high-dimensional abstract feature that can characterize the complex internal mechanism of the process, such as the nonlinear coupling relationship between the evolution of the material's microstructure and the macroscopic cutting force. Inputting this data into the final recommendation model for training enables the model to learn the mapping law between process parameters and final processing quality more accurately.

[0031] This invention performs statistical preprocessing on the integrated process data through standardization and principal component analysis, which can efficiently remove redundant information and random noise, providing a low-dimensional, decorrelation-free clean data space for subsequent nonlinear learning. This significantly improves the training stability of the subsequent model. Through the autoencoder, there is no need to predefine the mechanism model. It learns autonomously and extracts potential effective representations in a data-driven manner. The feature process data has high information density and strong discriminative power, effectively solving the black box problem of process mechanism and laying a high-quality data foundation for subsequent intelligent recommendation models.

[0032] S3. Perform density clustering on the feature process data to obtain cluster label vectors. Based on the cluster label vectors, perform isolated forest anomaly processing and anomaly cluster integration to generate clustered process data. Existing technologies usually rely solely on density clustering. Although it can discover clusters of arbitrary shapes, it is sensitive to global parameters. When faced with high feature dimensions and uneven density distribution of process data, it is difficult to set parameters. It is easy to misjudge sparse but valuable innovative process points as noise. Furthermore, when dealing with special cases such as special material processing or new process verification, there are often both sparse but reasonable and real interference or erroneous data in the data. Due to the lack of fine-grained ability to distinguish between sparsity and anomalies, existing technologies are prone to filtering out valuable features or causing anomalies to contaminate the clustering results. To solve the above problems, the specific steps are as follows: S31. Based on the feature process data, obtain the neighborhood parameters and minimum number of points through parameter estimation. In this step, for each input data sample point, calculate its Euclidean distance to all other sample points in the feature process dataset. The Euclidean distance is calculated by subtracting the values ​​of the two sample points in each feature dimension, squaring the difference, summing the squared differences in all feature dimensions, and finally taking the square root of the sum to obtain a distance value representing the overall difference between the two points. For each sample point, arrange its distance to all other points in ascending order and take the k-th position, where k is usually a small integer, such as the distance value of the feature dimension value, which is called the k-nearest neighbor distance of the point. Sort the k-nearest neighbor distances of all sample points and draw an increasing curve, i.e., the k-distance graph. In this figure, the inflection point where the slope of the curve changes significantly is identified. The distance value on the ordinate corresponding to this inflection point is selected as the neighborhood radius parameter. This inflection point means that the number of sample points with a distance less than this value will increase significantly, forming a boundary of density change. At the same time, the minimum number of points parameter is calculated, and its value is set to the dimension of the feature process dataset, that is, the number of features contained in each sample point multiplied by two. For example, if each sample has twenty features, then the minimum number of points is forty. This multiplication setting is based on experience and aims to ensure that the neighborhood of a core point contains enough points to form a statistically significant cluster structure. It is also related to the complexity of the data itself, that is, the dimension. The two parameters of neighborhood radius and minimum number of points obtained are the basis for the subsequent density clustering algorithm to correctly identify different density regions in the data.

[0033] S32. Based on feature process data, neighborhood parameters, and minimum number of points, perform core point detection and generate a core point set. In this step, first select a point to be judged and calculate the Euclidean distance from this point to all other points in the dataset. The Euclidean distance is calculated by subtracting the values ​​of the two points in each feature dimension to obtain the difference in each dimension, squaring each difference, summing the squared results in all dimensions, and finally taking the square root of the sum to obtain the final distance value. Next, based on all the distance values ​​calculated in the previous step, count how many other points are less than or equal to the given neighborhood radius parameter. Specifically, compare each calculated distance value with the neighborhood radius parameter. If the distance value is less than or equal to the radius, the point corresponding to that distance is included in the neighborhood. After completing the statistics, all neighboring points in a circular or high-dimensional spherical region centered on the current point and bounded by the neighborhood radius are obtained. The total number of these points is called the neighborhood density. Then, the neighborhood density is compared with the minimum number of points input. If the neighborhood density is greater than or equal to the minimum number of points, the current point is determined to be a core point and added to the core point set. Otherwise, the current point is temporarily marked as a non-core point. This process repeats the distance calculation, neighborhood statistics and comparison judgment operation for each point in the dataset. The final core point set is obtained, where each point represents a high-density region center in the feature space. In intelligent recommendation of metal processing parameters, these core points correspond to high-quality process modes that have been verified multiple times after feature extraction, have stable and reproducible parameter combinations, such as high-efficiency milling, high-surface-quality turning or the optimal heat treatment window for a specific alloy. Identifying these core points is the key first step for successful clustering. They will serve as anchor points, connecting and absorbing boundary points in their neighborhood through subsequent steps, thereby completely outlining different process clusters.

[0034] S33. Perform density clustering on the core point set, feature process data, and domain parameters to obtain cluster label vectors. In this step, first, all sample points are initially marked as unvisited. Then, the core point set is traversed. For each core point that has not yet been assigned to any cluster, the algorithm takes it as the starting point of a new cluster and performs cluster expansion. The expansion process adopts iterative search: First, the core point itself and all its neighboring points within the neighborhood radius are obtained by calculating the Euclidean distance between the two points and determining whether it is less than or equal to the radius parameter, and then added to the current cluster. The calculation of the Euclidean distance is as described above, involving subtraction, squaring, summation, and square root of corresponding feature values. Then, these newly added points are checked. If there are other core points among them, all unvisited neighboring points of these new core points are also added to the current cluster. This process is repeated continuously, checking each newly added core point and merging the points in its neighborhood until there are no more new points that can be added to the current cluster. At this time, a complete cluster is formed, and all its members are assigned the same cluster number label. After constructing a cluster, the algorithm continues to select the next unvisited core point as the starting point and repeats the above expansion process to form a second and third cluster, until all core points have been visited and processed. Finally, all sample points that do not belong to any cluster, i.e., neither core points nor included in the neighborhood expansion of any core point, are marked as noise points and assigned a special noise label, usually represented by negative one. The output result is a cluster label vector, which clearly indicates the process cluster class to which each sample point belongs or its status as noise. In the intelligent recommendation of metal processing parameters, the sample points after deep feature extraction are automatically divided into several clusters with high internal density and mutual separation. Each cluster corresponds to a processing mode with similar internal rules. For example, different clusters may represent stable and reproducible high-quality parameter combination areas such as high speed and small depth of cut for high surface finish, large depth of cut for low speed and high efficiency removal, or heat treatment under specific temperature control range. The intelligent recommendation mapping relationship from processing requirements, such as target surface finish, to the corresponding optimal parameter cluster is established, providing a clear and reliable data structure foundation.

[0035] S34. Perform isolated forest processing based on feature process data to obtain a set of isolated trees. In this step, a portion of the samples is randomly selected from the entire feature process dataset, for example, a fixed proportion of the original data volume, as the training subset for constructing a single isolated tree. For this subset, the recursive process for constructing an isolated tree is as follows: At the current node of the tree, first, a feature is randomly selected from all process feature dimensions. For example, it may be randomly selected between a higher-order feature one representing thermal effects and a higher-order feature two representing nonlinear cutting forces. Then, within all values ​​of the current node data for this feature, a splitting value is randomly selected. This value lies between the minimum and maximum values ​​of the feature in the data of this node. Then, based on... This splitting value divides the sample data of the current node into two groups: one group consists of samples whose feature value is less than the splitting value, and the other group consists of samples whose feature value is greater than or equal to the splitting value. This process is recursively repeated on the two newly created child nodes: in each child node, features and splitting values ​​are randomly selected again to divide the data. There are two termination conditions for the recursion: one is that the height of the tree reaches the preset maximum limit; the other is that the current node has only one sample data, or all samples have the same value on all features of the node, and cannot be further divided. At this time, the sample point is isolated in a leaf node of the tree. By repeating the above process a large number of times, a specified number of isolated trees are constructed. These trees together form an isolated forest, and the final set of isolated trees is obtained. The essence of an isolated tree ensemble is a model that integrates multiple random binary tree structures. In intelligent recommendation of metal processing parameters, since each tree isolates samples by randomly selecting features and segmentation values, abnormal process states that are significantly different from the mainstream patterns, such as distortion features caused by momentary sensor failures or rare parameter combinations caused by extreme improper operations, are quickly isolated in most trees due to their unique characteristics. They can usually be separated into a leaf node with only a short path length. Conversely, normal, highly clustered process patterns require more segmentation steps to be distinguished. Therefore, this isolated tree ensemble provides the computational foundation for the next step of calculating the abnormal score of each sample point, thereby accurately identifying potential abnormal or noisy process data, ensuring the purity and reliability of the dataset on which subsequent clustering analysis and intelligent recommendation depend.

[0036] S35. Path length processing is performed on the isolated tree set and feature process data to obtain the anomaly score vector. In this step, for the current sample point, the algorithm first traverses each isolated tree in the isolated forest. In each tree, starting from the root node, based on the pre-selected feature dimension and its segmentation threshold, it determines whether the current sample point's value on that feature is less than or greater than or equal to the threshold, thus deciding whether it enters the left or right subtree. This process is repeated layer by layer. Each time a tree node is traversed, the path length count is incremented by one until the sample point finally reaches a leaf node. The total number of nodes traversed from the root node to the leaf node in that tree is recorded, which is the path length of the sample point in that tree. After traversing all trees, the path lengths of the sample point in each tree are summed, and then divided by the total number of trees in the forest to obtain the average path length of the sample point. This average value reflects the typical number of segmentation steps required to isolate the point. Next, the calculation... A standardization constant related to the number of samples in the dataset is calculated using a harmonic series: First, the harmonic series of the total number of samples minus one is calculated by taking the reciprocals of all integers from one to the total number of samples minus one, summing them, multiplying this sum by two, and finally subtracting the quotient of twice the total number of samples minus one from the result. This gives the standardization constant. Then, the anomaly score is calculated: the average path length is divided by this standardization constant to obtain a ratio. Then, two is raised to the power of this ratio, i.e., using two as the base and the negative of the ratio as the exponent. The result is the anomaly score for that sample point. The closer the score is to one, the shorter the average path and the easier it is to isolate, indicating a higher probability of an anomaly. The closer the score is to zero, the longer the path and the more it belongs to a normal, dense group. The resulting anomaly score vector provides a continuous anomaly measure between zero and one for each process data sample.

[0037] S36. Perform anomaly clustering integration on the anomaly score vector and cluster label vector to generate clustering process data. In this step, first traverse all sample points in the cluster label vector. For samples that have been marked as noise points by density clustering, i.e., the label is negative one, check their corresponding anomaly score and compare this anomaly score with a preset anomaly threshold, usually set to 0.5. If the anomaly score of the sample is greater than the threshold, it is determined to be a highly anomalous outlier and its label is updated to a new, independent anomaly category label, for example, the label is set to negative two. If the anomaly score of the sample is less than or equal to the threshold, it means that although it is not covered by the core point density, it does not show strong anomalous characteristics and may belong to the edge point of a certain cluster. For such points, the algorithm will perform label reassignment: calculate the distance from the point to each sample point in all non-noise clusters. The distance calculation uses the Euclidean distance formula, which involves subtracting, squaring, summing and then taking the square root of each feature value, which will not be elaborated here. The process involves identifying the nearest non-noise sample point and assigning its cluster label to the current point, thus placing it into a nearby, reasonable process cluster. This process ensures that valuable edge process parameter combinations are not incorrectly discarded. For sample points already divided into clusters in density clustering (i.e., points with positive integer labels), their labels are typically retained and used directly as the final labels. The resulting clustered process dataset not only contains the final process cluster number of each sample point but also includes information to determine whether it is a definitive outlier. The clustered process data includes process parameters and performance data. In intelligent recommendation of metal processing process parameters, this is achieved by introducing... The anomaly score, an external metric, provides fine-grained correction for potential misjudgments in density clustering, resulting in more homogeneous and pure process clusters. For example, pseudo-samples that may be mixed into the high-efficiency, low-wear milling parameter cluster due to momentary interference during data acquisition will be identified as anomalies and removed. Meanwhile, some experimental process points that are not dense but have reasonable parameters are correctly assigned to neighboring clusters. This ensures that the process features mined and constructed based on cluster statistics, such as calculating the cluster centers and parameter distributions, have extremely high reliability and representativeness. Consequently, the anomaly classification results are of higher quality, higher accuracy, and greater stability.

[0038] This invention reduces subjective dependence by using automated methods such as distance maps, enabling adaptive discovery of density clusters of arbitrary shapes and sizes in the feature space. This aligns with the complex distribution patterns that may result from combinations of process parameters. Simultaneously, using the high-dimensional features extracted earlier as input, the isolated forest employs a random partitioning strategy to evaluate the segregability of each sample point from another dimension, thereby calculating a density-independent global anomaly score. Based on the initial clustering result labels, the anomaly score is used for fine-tuning, achieving a balance between noise filtering and edge information preservation. This significantly improves the purity and completeness of the clustering results. When facing unconventional scenarios such as special materials, this combined method demonstrates strong robustness, resulting in higher quality clustered process data and providing high-quality, highly stable data for subsequent steps.

[0039] S4. Based on the clustered process data, reinforcement learning actions are obtained through policy reasoning. Multi-objective optimization is performed based on the clustered process data and reinforcement learning actions to generate a recommended set of process parameters. Existing technologies simply use reinforcement learning based on historical experience. Although it can make recommendations using known data, its exploration ability is limited. When facing new process scenarios with sparse data, it is easy to get stuck in local optima. Traditional Bayesian optimization search efficiency is low, and when facing high-dimensional and complex process parameter spaces, the computational cost is high, and the model is difficult to build accurately. In practice, it is easy to encounter special high-risk scenarios such as using special materials for first-piece processing or breaking through process limits. At this time, it is often necessary to find feasible or even optimal parameters within a very limited number of trial and error opportunities. Existing technologies either recommend mediocre solutions due to being too conservative or lead to uncontrollable trial and error costs due to blind exploration. They cannot achieve safe and efficient parameter optimization with limited resources. To solve the above problems, the specific steps are as follows: S41. Extract context features based on clustered process data to obtain clustered context vectors. In this step, firstly, based on the final cluster labels, all sample points belonging to the current target cluster are selected from the clustered process dataset. These points represent a group of processing instances that are close to each other in the feature space and have similar process patterns. Then, for all samples within this cluster, feature centers are calculated. The calculation is performed independently for each feature dimension: for a specific process feature dimension, such as a deep nonlinear feature reflecting cutting stability, the specific values ​​of all sample points within the cluster on this feature are extracted, and then these... The numerical values ​​are summed, that is, the feature values ​​of the first sample, the second sample, and so on, up to the last sample, and a sum is obtained. Then, this sum is divided by the total number of sample points in the cluster. The result of this division operation represents the average level or typical value of the process feature in the process mode represented by the current cluster. The above extraction, summation and division by the number of samples are repeated for all other feature dimensions in the feature process dataset. Finally, the average values ​​calculated on all feature dimensions are combined into a new vector according to the feature order, which is the clustering context vector of the cluster. In intelligent recommendation of metal processing parameters, each clustering context vector is essentially a digital fingerprint of a stable and reproducible combination of high-quality processing parameters. For example, a cluster may represent an efficient rough milling process for a certain type of stainless steel. The values ​​in its context vector respectively characterize the comprehensive typical state of key parameters such as rotational speed, feed rate, and depth of cut under this process mode after noise reduction and abstraction. This vector not only contains the average level of the parameters, but more importantly, it reflects the optimized matching relationship between parameters based on depth features. The clustering context vector constitutes the core process knowledge base of the subsequent system.

[0040] S42. Based on the clustering context vector, a reinforcement learning state is generated through state encoding. In this step, the input clustering context vector is first used as the initial data for the network input layer. Each value in this vector corresponds to a process feature after deep extraction and cluster averaging, such as a nonlinear factor characterizing the cutting resistance of a specific material or a comprehensive index reflecting thermal management efficiency. Each neuron in the first layer of the network performs a calculation: multiplying the set of weight parameters corresponding to the neuron by the value at each position in the input vector, then summing all these products, and finally adding a bias parameter value associated with the neuron. This process is called weighted summation. The summation result is then input into a nonlinear activation function, such as a linear rectifier function. In this function, if the input value is greater than zero, the output is unchanged; if it is less than or equal to zero, the output is zero. This introduces non-linear transformation capability. This calculation is performed independently on each neuron in this layer and is passed to the next layer of the network. The next layer also contains multiple neurons, and each neuron performs the same operation again: multiply each value of the previous layer's output by a new set of weight parameters corresponding to this neuron, sum them, add a bias, and then pass them through the activation function. This process is passed and transformed layer by layer, usually through two or more layers. In the last layer, the vector dimension of the network output is designed to be a preset fixed length. This vector is the final state vector. The final reinforcement learning state vector is an abstract representation of the original process pattern that has been deeply re-encoded.

[0041] S43. Perform policy reasoning on the reinforcement learning state to obtain the reinforcement learning action. In this step, the input state vector is first fed into the first layer of the network. Each neuron in this layer stores a set of adjustable weight parameters and a bias parameter. The calculation of the neuron is divided into two steps: The first step is linear transformation, which multiplies each value of the state vector with each value of the corresponding weight parameter of the neuron, and then sums all these products, and finally adds the bias parameter value of the neuron. The second step is nonlinear activation, which inputs the weighted sum obtained in the previous step into an activation function, such as a linear rectified function. This function stipulates that if the input value is greater than zero, the output is the value itself, and if the input value is less than or equal to zero, the output is zero. This calculation process is processed in parallel on all neurons in the first layer to generate a new set of feature values. This new set of feature values ​​is passed to the subsequent layers of the network as an intermediate result. Each subsequent layer repeats the same operation as the first layer: receiving the output vector of the previous layer, each neuron multiplies each value of the vector of the previous layer with its own stored weight parameters, sums them and adds the bias, and then generates a new output through the activation function. After several layers of such progressively nonlinear transformations, the data is passed to the last layer of the network, the output layer. The number and design of neurons in the output layer correspond to the dimensions of process parameter adjustment. This layer also performs linear calculations of weighted summation and bias addition, and its output is the final action vector. Usually, this vector is then passed through a scaling function, such as the hyperbolic tangent function, to map the value of each dimension to a preset, physically meaningful adjustment range. The resulting reinforcement learning action vector has each dimension directly corresponding to the adjustment amount or direction of a key process parameter. In intelligent recommendation of metal processing parameters, for example, the action vector may indicate: under the current state of efficient milling of a certain alloy, it is recommended to increase the spindle speed by a certain specific value and decrease the feed per tooth by a certain specific value. This action is generated based on the policy network's learning of the complex relationship between massive historical process data and processing results, and can guide the processing system to optimize towards better performance indicators, such as higher material removal rate or longer tool life.

[0042] S44. Extract historical process parameters and performance data from historical clustered process data, and perform Gaussian regression on the historical process parameters and performance data to obtain a Gaussian processing model. In this step, for any two process parameter points, calculate the difference between them. Subtract the value of the corresponding dimension of the second parameter point from the value of each dimension of the first parameter point to obtain the difference in each dimension. Then, square each difference. Then, sum the squared values ​​of all dimensions to obtain the squared distance reflecting the overall difference between the two points. Next, divide this squared distance by twice the square of a parameter called the length scale parameter. This length scale parameter controls the rate at which performance similarity decays with parameter distance and is a key hyperparameter of the model. Then, take the negative value of the quotient obtained from the above division. Finally, perform an exponential operation with the natural constant as the base of this negative value. The result is the covariance function value between the two parameter points. This value is between zero and one. It is one when the two parameters are exactly the same and approaches zero as the parameter difference increases, indicating that the expected performance of the two points is less correlated. Based on the covariance calculation between all pairwise historical data points, a complete covariance matrix can be constructed. The Gaussian process model assumes that the processing performance index follows a Gaussian process with zero as the mean function and the covariance function calculated above as the core. By substituting historical performance data into this model for fitting, the parameters of the model can be determined. The resulting Gaussian process model can provide a predicted processing performance distribution for any new and unseen set of process parameters, including a most likely predicted mean and a predicted variance representing the confidence level. In intelligent recommendation of metal processing parameters, this model simulates expensive and time-consuming real processing experiments, enabling rapid evaluation of the expected effects and reliability of a large number of candidate parameter combinations in virtual space, providing accurate processing data for subsequent Bayesian optimization algorithms.

[0043] S45. Based on the Gaussian processing model and reinforcement learning state, perform expected improvement processing to obtain the acquisition function value. In this step, firstly, obtain the currently known best measured value of processing performance from historical data, which is denoted as the historical best performance value. For any new candidate process parameter combination to be evaluated, the Gaussian process model will give a probability distribution of its predicted performance. This distribution is described by a prediction mean and a prediction variance. The core idea of ​​calculating the acquisition function is to evaluate the expected improvement of the candidate point beyond the historical best performance value. The specific calculation is approximated by the Monte Carlo sampling method. The specific steps are as follows: First, perform a large number of random samples from the Gaussian process prediction distribution of the candidate point, that is, generate multiple sets of random performance prediction values ​​that conform to the normal distribution based on its prediction mean and variance. For each random predicted value obtained from sampling, the following operation is performed: subtract the historical best performance value from the random predicted value to obtain a difference. Then compare this difference with zero. If the difference is greater than zero, it means that the result of this random sampling is better than the historical best, so the difference is retained. If the difference is less than or equal to zero, it means that the result of this sampling is not better than the historical best, so the improvement is recorded as zero. Mathematically, this step is represented by taking the maximum value between the difference and zero. After completing the operation on all a large number of random samples, all results are summed, and then the sum is divided by the total number of samplings to obtain the average expected improvement. This average value is the acquisition function value of the candidate parameter point. The final acquisition function value is a non-negative real number, which comprehensively measures the expected benefit brought by the potential and uncertainty of the performance of this point to surpass the historical best.

[0044] S46. Based on the acquired function values ​​and reinforcement learning actions, a recommended set of process parameters is generated through multi-objective optimization. In this step, based on the current state and the actions suggested by reinforcement learning, the long-term expected return value corresponding to the set of actions is queried or calculated from the evaluation network of the action value function. The evaluation network of the action value function is a core deep neural network component in the reinforcement learning algorithm. It is specifically used to learn and predict the long-term, comprehensive expected processing performance return that can be brought about after executing a specific parameter adjustment action. This value quantifies the stable processing performance gain that can be brought about after executing the set of parameter adjustments. At the same time, for several specific candidate process parameters mapped by the action vector or derived in its neighborhood, such as different combinations of rotation speed and feed, their respective acquired function values ​​are obtained from the Bayesian optimization module. This value quantifies the expectation that trying the set of parameters may bring breakthrough improvements. Next, multi-objective fusion calculation is performed: the expected return value of the action obtained in the first step is multiplied by a preset utilization weight to obtain a weighted utilization score; the acquisition function value of the candidate parameter obtained in the second step is multiplied by a preset exploration weight to obtain a weighted exploration score. Both weights are decimals between zero and one, and their sum is one. Then, for each candidate parameter group, its corresponding utilization score and exploration score are added to obtain a comprehensive evaluation value. Finally, among all the evaluated candidate process parameter groups, comparisons and selections are made to find the parameter combination that maximizes the above comprehensive evaluation value. This parameter group is selected as the final recommendation. The final recommended process parameter dataset contains one or more sets of specific, executable machining instructions, such as optimal spindle speed, feed rate, and depth of cut.

[0045] This invention constructs reinforcement learning states based on context vectors extracted from early clustering analysis, enhancing the baseline reliability and security of recommendations. A Bayesian optimization module based on Gaussian processes performs probability modeling and uncertainty quantification for unknown regions. Multi-objective optimization fusion uses a weighted fusion formula to provide a comprehensive score for each recommendation. When facing black-box scenarios such as first-piece processing of special materials or breakthroughs in process limits, this combined method demonstrates excellent adaptability and robustness, avoiding catastrophic failures while actively seeking optimal solutions, thus generating optimal process parameter recommendations efficiently and reliably.

[0046] S5. Verify and fine-tune the recommended process parameter set to generate optimized process parameter data. Existing technologies usually rely on actual experiments for parameter verification, which is costly, time-consuming and risky. Performance evaluation often depends on a single indicator, lacks comprehensive consideration, and is difficult to achieve accurate and automated parameter optimization. To solve the above problems, the specific implementation steps are as follows: S51. Based on the recommended set of process parameters, construct a digital twin and generate simulated performance data. In this step, firstly, based on the input parameters, such as cutting speed, and combined with the hardness, toughness and other properties of the workpiece material, calculate the instantaneous cutting force through the material removal mechanics equation. This calculation involves substituting multiple variables such as material shear strength and tool geometry into the formula for iterative solution. Then, based on the calculated cutting force and friction coefficient, use the law of conservation of energy to deduce the heat generated in the processing area, and simulate the diffusion process of heat in the workpiece, tool and chips through the partial differential equation of heat conduction to predict the spatiotemporal distribution of temperature. Subsequently, based on the predicted temperature field and stress field, the model will call the material phase transformation and thermo-mechanical coupling model to evaluate the residual stress, microstructure changes and wear rate of the tool rake face and flank face that may be generated on the workpiece surface. The final model integrates the simulation results of all the above mechanical, thermal, and materials science states, and calculates quantifiable performance indicators through a series of predefined evaluation functions. For example, the processing quality score is obtained by substituting the simulated surface roughness prediction value, dimensional tolerance deviation prediction value, etc. into a weighted summation formula, and the processing efficiency can be obtained by dividing the material removal rate prediction value by the energy consumption prediction value per unit time to get a comprehensive score. All these calculations are completed in the computer. The final simulation performance data provides a detailed virtual test report for each set of input recommended process parameters. In the intelligent recommendation of metal processing process parameters, the feasibility and potential effects of the parameters are evaluated in advance before they are put into expensive actual production or physical testing, which greatly reduces the cost and risk of trial and error.

[0047] S52. Based on the simulated performance data, the evaluation results are obtained through key indicator processing. This step first extracts two core performance indicators from the input simulated performance data: the first is the processing quality score, which integrates simulated predicted surface roughness, dimensional accuracy, residual stress, and other quality-related factors. It is usually normalized to a score between zero and one hundred; a higher score indicates better processing quality. The second is the processing efficiency value, which integrates predicted material removal rate, unit product energy consumption, or production cycle time, and is also usually normalized to a score between zero and one hundred; a higher score indicates better efficiency. Next, a weighted comprehensive calculation is performed: two preset weighting coefficients are called. These correspond to the relative importance of quality and efficiency, respectively. Both weighting coefficients are decimals between zero and one, and their sum is strictly equal to one. This ensures the balance of the evaluation system. The first step in the calculation is to multiply the extracted quality score by the corresponding quality weighting coefficient to obtain a weighted quality contribution value. The second step is to multiply the extracted efficiency value by the corresponding efficiency weighting coefficient to obtain a weighted efficiency contribution value. Then, these two weighted contribution values ​​are added together, and the sum is the evaluation result of the recommended process parameters. The final evaluation result is a single numerical value that quantitatively reflects the comprehensive expected performance of the parameter scheme after weighing specific production goals.

[0048] S53. Fine-tune the parameters of the evaluation results to obtain the optimal process parameter data. In this step, the input evaluation results are compared with a preset performance qualification threshold, for example, a performance qualification threshold of 0.8. If the evaluation result is less than the threshold, it indicates that the expected comprehensive performance of the current recommended parameters has not reached the optimal standard, and the parameter adjustment process will be initiated. The core of the adjustment is to calculate the gradient of the evaluation result relative to each dimension of the recommended process parameters. This gradient calculation is completed by the automatic differentiation function inside the digital twin simulation system. It can accurately give the slight increase or decrease of each process parameter such as spindle speed, feed rate, etc. Then, these calculated gradient vectors are multiplied by a preset value less than one. A positive learning rate, such as 0.02, is used to control the step size of each adjustment, preventing excessive parameter changes. Then, for each parameter in the original recommended process parameter dataset, the product of its corresponding gradient component and the learning rate is added. This addition operation means that the original parameters are increased or decreased in small, purposeful ways along the direction that most quickly improves the evaluation results. If the evaluation result is greater than or equal to the threshold, it is determined that the current recommended parameters meet the requirements, and no adjustment is made. This result is directly used as the final output, and the optimized process parameter data is obtained. This is the result of the original intelligent recommendation scheme after a round of performance feedback and gradient optimization based on physical simulation and data-driven methods.

[0049] This invention achieves improved efficiency and accuracy through virtual verification and optimization using digital twin simulation, comprehensive evaluation, and gradient feedback fine-tuning. It utilizes a high-fidelity physical model to safely, quickly, and cost-effectively predict processing results in the digital space, avoiding material waste, equipment damage, and production delay risks. Through comprehensive evaluation of multiple indicators with configurable weights, it flexibly adapts to different production optimization goals. The introduction of a gradient-based automatic feedback adjustment mechanism enables automatic and accurate parameter correction based on simulation performance, significantly improving the reliability and optimality of the final recommended parameters.

[0050] Example 2: Because existing technologies are prone to getting trapped in local optima, have high computational costs, and are difficult to construct accurately, they struggle to achieve safe and efficient parameter optimization with limited resources. Please refer to [link to relevant documentation]. Figure 2 The diagram shown is a structural block diagram of the intelligent recommendation system for metal processing parameters based on big data provided in this embodiment. The system includes an integration module, a feature module, a clustering module, a recommendation parameter module, and an optimization parameter module.

[0051] The integration module is used to collect the raw process data stream of the metal processing system and generate integrated process data based on the raw process data stream through integration processing. The feature module is used to perform covariance decomposition based on integrated process data to obtain eigenvalue vectors and eigenvector matrices, and then perform dimensionality reduction encoding on the eigenvalue vectors and eigenvector matrices to obtain feature process data. The clustering module is used to perform density clustering on the feature process data to obtain cluster label vectors. Based on the cluster label vectors, isolated forest anomaly processing and anomaly cluster integration are performed to generate clustered process data. The recommended parameter module is used to obtain reinforcement learning actions by processing the clustered process data through policy reasoning, and to perform multi-objective optimization based on the clustered process data and reinforcement learning actions to generate a recommended process parameter set. The optimal parameter module is used to verify and fine-tune the recommended process parameter set to generate optimal process parameter data.

[0052] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0053] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for intelligent recommendation of metal processing parameters based on big data, characterized in that, The steps of this method are as follows: acquire the raw process data stream of the metal processing system, and generate integrated process data based on the raw process data stream through integration processing; Based on the integrated process data, covariance decomposition is performed to obtain eigenvalue vectors and eigenvector matrices. The eigenvalue vectors and eigenvector matrices are then subjected to dimensionality reduction encoding to obtain characteristic process data. Density clustering is performed on the characteristic process data to obtain cluster label vectors. Based on the cluster label vectors, isolated forest anomaly processing and anomaly cluster integration are performed to generate clustered process data. Based on the clustered process data, reinforcement learning actions are obtained through policy reasoning. Multi-objective optimization is then performed based on the clustered process data and reinforcement learning actions to generate a recommended set of process parameters. The recommended process parameter set is verified and fine-tuned to generate optimal process parameter data.

2. The intelligent recommendation method for metal processing parameters based on big data according to claim 1, characterized in that, Based on the original process data stream, the data is integrated and processed, including: multi-source data processing based on the original process data stream to generate the original process dataset; The cleaning process dataset is obtained by processing outliers and missing values ​​based on the original process dataset; The cleaning process datasets were standardized and merged to obtain integrated process data.

3. The intelligent recommendation method for metal processing parameters based on big data according to claim 1, characterized in that, Covariance decomposition based on integrated process data includes: standardization based on integrated process data to obtain a standardized process dataset. Covariance processing is performed on the standardized process dataset to generate a covariance matrix; The covariance matrix is ​​subjected to eigenvalue decomposition to generate eigenvalue vectors and eigenvector matrices.

4. The intelligent recommendation method for metal processing parameters based on big data according to claim 1, characterized in that, Dimensionality reduction encoding of eigenvalue vectors and eigenvector matrices includes: dimensionality reduction projection of the standardized process dataset, eigenvalue vectors, and eigenvector matrices to obtain dimensionality-reduced process data; Encoding feature vectors are generated by performing encoding feature processing on dimensionality-reduced process data. Feature process data is obtained by performing feature selection processing based on the encoded feature vector.

5. The intelligent recommendation method for metal processing parameters based on big data according to claim 1, characterized in that, Density clustering of characteristic process data includes: obtaining neighborhood parameters and minimum number of points by parameter estimation based on the characteristic process data; Core point detection is performed based on feature process data, domain parameters, and minimum number of points to generate a core point set. Density clustering is performed on the core point set, feature process data, and domain parameters to obtain cluster label vectors.

6. The intelligent recommendation method for metal processing parameters based on big data according to claim 1, characterized in that, Anomaly processing and anomaly clustering integration of isolated forests based on clustering label vectors include: processing isolated forests based on feature process data to obtain a set of isolated trees; Path length processing is performed on the isolated tree set and feature process data to obtain the anomaly score vector; Anomaly clustering is performed on the outlier score vector and cluster label vector to generate clustering process data.

7. The intelligent recommendation method for metal processing parameters based on big data according to claim 1, characterized in that, The clustering process data is processed through strategy reasoning, including: extracting contextual features based on the clustering process data to obtain a clustering context vector; Reinforcement learning states are generated by processing the clustering context vectors through state encoding. The reinforcement learning state is processed by policy reasoning to obtain the reinforcement learning action.

8. The intelligent recommendation method for metal processing parameters based on big data according to claim 1, characterized in that, Multi-objective optimization is performed based on clustering process data and reinforcement learning actions, including: extracting historical process parameters and performance data from historical clustering process data, performing Gaussian regression on the historical process parameters and performance data, and obtaining a Gaussian processing model; Based on the Gaussian processing model and reinforcement learning state, the desired improvement processing is performed to obtain the acquisition function value; Based on the collected function values ​​and reinforcement learning actions, a set of recommended process parameters is generated through multi-objective optimization.

9. The intelligent recommendation method for metal processing parameters based on big data according to claim 1, characterized in that, The recommended process parameter set is validated and fine-tuned, including: constructing a digital twin based on the recommended process parameter set to generate simulated performance data; The evaluation results are obtained by processing the simulated performance data through key indicators; The evaluation results are fine-tuned to obtain the optimal process parameter data.

10. A system for applying the intelligent recommendation method for metal processing parameters based on big data as described in any one of claims 1-9, characterized in that, The system includes: The integration module is used to collect the raw process data stream of the metal processing system and generate integrated process data based on the raw process data stream through integration processing. The feature module is used to perform covariance decomposition based on integrated process data to obtain eigenvalue vectors and eigenvector matrices, and then perform dimensionality reduction encoding on the eigenvalue vectors and eigenvector matrices to obtain feature process data. The clustering module is used to perform density clustering on the feature process data to obtain cluster label vectors. Based on the cluster label vectors, isolated forest anomaly processing and anomaly cluster integration are performed to generate clustered process data. The recommended parameter module is used to obtain reinforcement learning actions by processing the clustered process data through policy reasoning, and to perform multi-objective optimization based on the clustered process data and reinforcement learning actions to generate a recommended process parameter set. The optimal parameter module is used to verify and fine-tune the recommended process parameter set to generate optimal process parameter data.

Images