Multi-axis numerical control machining parameter adaptive optimization method for precise automobile components

Abstract

The present application relates to the technical field of numerical control machining, in particular to a multi-axis numerical control machining parameter adaptive optimization method for automobile precision components. The method comprises the following steps: obtaining historical data containing machining parameters, geometric feature vectors, machining quality data and sensor time series; constructing a hierarchical latent variable model, mapping the geometric feature vectors to a first latent space through manifold learning, mapping the machining parameters and geometric latent variables to a second latent space through a Gaussian process, determining the kernel length scale by correlating the tool wear state evaluation value and the geometric latent variables, and mapping the parent latent variables of the mixed Gaussian prior to the machining quality data; inputting the geometric feature vectors of the to-be-machined area, and optimizing the machining parameters through the weighted sum of the risk quantile and the expected value of the posterior predictive distribution and back propagation. The scheme of the present application uniformly describes the cutting nonlinear response and the tool state, and improves the machining quality consistency and the parameter optimization reliability.

Description

Technical Field

[0001] This invention relates to the field of CNC machining technology, and more specifically to an adaptive optimization method for multi-axis CNC machining parameters of automotive precision components. Background Technology

[0002] With the rapid development of the modern automotive industry, automotive components are facing higher demands for machining accuracy, surface quality, and production efficiency. CNC machining, as a crucial link in automotive component manufacturing, directly impacts the safety performance and service life of vehicles. In actual CNC machining processes, machining quality is influenced by multiple factors, including the CNC machining parameters set by the machine tool, the geometric characteristics of the component itself, and the changing operating conditions over time. To achieve intelligent manufacturing, it is necessary to fully utilize historical machining data accumulated during production, including machining parameters, geometric features, time-series signals obtained from real-time sensor monitoring, and machining quality data. Due to the significant nonlinear characteristics of CNC machining systems, there is a complex coupling relationship between the spatial constraints of the component's geometric surface and the temporal evolution characteristics of the machining process. How to uncover the hidden patterns in multi-source heterogeneous data and establish an effective mapping relationship from input features to output machining quality has become a crucial problem to be solved in the field of high-end CNC manufacturing.

[0003] In the field of CNC machining parameter optimization for automotive components, existing prediction and optimization models typically model surface geometric constraints and dynamic characteristics of the machining process separately when dealing with high-dimensional geometric features and time-series sensor data. This makes it difficult to form a multi-level, interconnected feature representation system. This separate processing approach fails to fully characterize the correlation between tool wear evolution, local surface features, and changes in machining state, leading to biases in the model's description of the evolution of the actual machining process. When performing back-optimization of parameters for unknown areas to be machined, existing optimization objectives are usually relatively singular, making it difficult to balance the desired machining quality with potential fluctuation risks, and also difficult to adaptively adjust optimization weights according to the machining difficulty corresponding to different local geometric features. Therefore, in complex and variable actual machining environments, the optimized CNC machining parameters often lack adaptability and stability, affecting the consistency of automotive component machining quality.

[0004] Chinese patent document CN120871745B discloses an adaptive machining parameter optimization method and system for CNC machine tools. It detects tool wear, predicts the wear state in conjunction with the timing of the machining process, and adaptively optimizes CNC machining parameters accordingly to maintain stable machining quality. However, the optimization in this patent primarily focuses on changes in tool wear, offering limited ability to uniformly depict the relationship between workpiece surface geometry and cutting process dynamics. In machining scenarios with diverse surface shapes or complex geometric transitions, it struggles to make the prediction model adaptively adjust to the changing working conditions, resulting in insufficient adaptability of parameter optimization. Furthermore, its optimization orientation prioritizes the average machining quality level, lacking targeted control over localized extreme degradation risks and failing to simultaneously address the synergistic control of average machining accuracy and extreme deviation risks.

[0005] Chinese patent application CN121934382A discloses an adaptive optimization method and system for CNC machining process parameters based on reinforcement learning. This method collects machining data from multiple sources, quantifies measurement uncertainties, and performs planning and risk assessment throughout the tool's entire lifecycle, thereby achieving adaptive optimization of process parameters. While this method incorporates uncertainty quantification and risk assessment into parameter optimization, its optimization relies on reinforcement learning and meta-learning frameworks, focusing more on iterative convergence of the overall strategy. It lacks fine-grained differentiation of local geometric complexity differences in different curved surface regions, making it difficult to dynamically adjust the optimization focus based on local morphology. In regions with abrupt geometric changes or rapidly changing working conditions, its ability to balance average machining accuracy and extreme risks remains insufficient, and the stability of prediction and optimization needs improvement. Summary of the Invention

[0006] To address the problems of fragmented modeling of geometric features and temporal operating conditions, insufficient correlation of tool wear, and difficulty in balancing average quality and extreme risks in existing technologies, this invention proposes an adaptive optimization method for multi-axis CNC machining parameters of automotive precision components. The method includes the following steps: acquiring historical machining data, which includes input CNC machining parameters and input component geometric feature vectors, as well as a time series containing output machining quality data and output real-time sensor data; constructing a hierarchical latent variable model to establish a mapping relationship from input to output, wherein the process of constructing the hierarchical latent variable model includes mapping the input component geometric feature vectors to a first latent space using a manifold learning algorithm, and combining the input CNC machining parameters with the geometric latent variables in the first latent space, mapping them to a Gaussian process. In the second latent space, the Gaussian process is trained using the time series of the output real-time sensor data and determines the tool wear state assessment value based on tool usage information. The kernel length scale of the covariance function of the Gaussian process is determined by associating the tool wear state assessment value with the geometric latent variable. Parent latent variables with a prior Gaussian mixture distribution are constructed in the first and second latent spaces, and a mapping between the parent latent variables and the output machining quality data is established. The geometric feature vector of the area to be machined is input into the trained hierarchical latent variable model, and the CNC machining parameters are optimized through backpropagation. The optimization objective function is the weighted sum of the risk quantiles and expected values ​​of the posterior prediction distribution of machining quality, and the weights of the weighted sum are adjusted according to the representation of the geometric feature vector in the first latent space.

[0007] By mapping the component geometric feature vectors to the first latent space through manifold learning to preserve the inherent geometric constraints of local surfaces, and then combining CNC machining parameters with geometric latent variables and mapping them to the second latent space through a Gaussian process, the nonlinear response of the cutting process is characterized. Simultaneously, the Gaussian process is trained using real-time sensor time series, and the kernel length scale of the covariance function is determined by associating the tool wear state assessment value with the geometric latent variables, avoiding prediction inaccuracies caused by using a fixed model under different tool states and surface morphologies. Prior parent latent variables with a mixed Gaussian distribution are constructed in the first and second latent spaces and mapped to machining quality data, establishing a stable mapping between multi-source heterogeneous inputs and quality outputs. Backpropagation optimization is performed using the weighted sum of the risk quantiles and expected values ​​of the posterior prediction distribution of machining quality as the optimization objective, and adjusting the weights according to the representation of geometric features in the first latent space. This avoids neglecting geometric differences and extreme degradation risks in single-parameter optimization, improving the adaptability and reliability of CNC machining parameter optimization.

[0008] Furthermore, acquiring the historical machining data includes: collecting the three-phase current signal of the CNC machine tool servo drive, calculating the absolute current deviation value of each sampling period to form the time series of the output real-time sensor data; extracting the normal vector angle and principal curvature of the surface of the component's three-dimensional model to generate the input component geometric feature vector; acquiring the numerical combination of feed rate and cutting depth as the input CNC machining parameters; and using a coordinate measuring machine to scan the machined area and calculate the average absolute deviation between the actual contour points and the theoretical positioning coordinates as the output machining quality data.

[0009] By collecting three-phase current signals from the servo drive of CNC machine tools and calculating the absolute deviation of the current in each sampling period to form a time series, geometric feature vectors are generated by extracting the normal vector angle and principal curvature of the surface of the component's three-dimensional model. The numerical combination of feed rate and depth of cut is used as machining parameters, and the average absolute deviation between the actual contour points and the theoretical positioning coordinates is calculated by scanning the machined area with a coordinate measuring machine as machining quality data. This allows cutting load, surface morphology, machining commands, and quality results to correspond to specific physical quantities, avoiding training bias caused by the mixing of data from different sources or inconsistencies in dimensions, and providing a data foundation with good targeting and consistency for the hierarchical latent variable model.

[0010] Furthermore, the manifold learning algorithm is the Isomap algorithm. The Isomap algorithm is used to reduce the dimensionality of the input component geometric feature vectors and map them to the first latent space to obtain geometric latent variables representing the geometric constraints of the local surface. The input CNC machining parameters and the geometric latent variables are tensor concatenated and used as the joint independent variable input of the Gaussian process and mapped to the second latent space. The time series of the output real-time sensor data is extracted by a long short-term memory network and then reduced in dimensionality to obtain a scalar process state index, which is used as the training dependent variable of the Gaussian process.

[0011] By employing the Isomap algorithm to reduce the dimensionality of the component's geometric feature vectors and mapping them to the first latent space, geometric latent variables representing the geometric constraints of the local surface are obtained. These latent variables are then concatenated with the CNC machining parameter tensor as joint independent variables of the Gaussian process. Simultaneously, the real-time sensor time series is extracted through a long short-term memory network and reduced to scalar process state indicators as training dependent variables. This avoids redundancy and noise interference caused by high-dimensional geometric features and the original time series directly participating in regression, allowing process dynamics to participate in modeling in a compact scalar form and improving the stability of the input-to-process state mapping.

[0012] Further, mapping to the first latent space using the Isomap algorithm includes: constructing a K-nearest neighbor graph based on the input component geometric feature vectors, calculating the Euclidean distance between nodes as initial edge weights; constructing a geodesic distance matrix by calculating the shortest path between any two nodes in the graph using the Dijkstra algorithm; performing a multidimensional scaling transformation on the geodesic distance matrix, calculating the low-dimensional embedding coordinates of each sample based on the top three eigenvalues ​​and their corresponding eigenvectors, and using the low-dimensional embedding coordinates as geometric latent variables mapped to the first latent space.

[0013] Furthermore, determining the tool wear condition assessment value includes: using a pre-established Taylor tool life equation. Calculate the current tool wear condition assessment value, where The current cutting speed, The theoretical tool life at the stated cutting speed is... and The life coefficient is pre-calibrated based on the tool material, workpiece material, and cutting conditions; the current cutting speed is calculated based on the machine tool spindle speed and tool diameter. , and by Calculate the theoretical tool life; read the current cumulative cutting time of the tool. ,calculate and will Limit the amplitude to the range of 0 to 1 to obtain the current tool wear status assessment value.

[0014] By using the pre-established Taylor tool life equation, the current cutting speed is calculated from the machine tool spindle speed and tool diameter to obtain the theoretical tool life. Then, the cumulative cutting time of the tool is read to calculate its ratio with the theoretical life and limit it to the range of zero to one to obtain the tool wear status assessment value. This transforms the degree of tool wear into a continuous index that changes monotonically with the cumulative cutting time and has bounded values, avoiding the neglect of tool wear or its out-of-bounds values ​​in the modeling process. This provides a stable and reliable input basis for the subsequent adjustment of the covariance function kernel length scale.

[0015] Furthermore, regarding the geometric latent variables Standardize the data; evaluate the tool wear condition. and Concatenate into a joint vector Calculate intermediate variables ,in For the preset weight vector, For preset bias; according to Calculate the scaling adjustment coefficient ,in and These are the preset lower and upper limits of the adjustment coefficients, both of which are positive numbers; the basic kernel length scale is adjusted by the scale adjustment coefficients. The current kernel length scale is obtained by multiplying, and the covariance matrix of the Gaussian process is updated using the current kernel length scale.

[0016] By standardizing the geometric latent variables and concatenating them with the tool wear condition assessment values ​​to form a joint vector, the linear intermediate variables are first calculated and then mapped using a scale adjustment coefficient formula with upper and lower limits. Finally, the basic kernel length scale is multiplied by this coefficient to obtain the current kernel length scale. This allows the kernel length scale to be smoothly adjusted with tool wear and surface geometry changes and is always constrained between preset upper and lower limits. This avoids the covariance matrix from becoming unstable due to excessively large or small adjustment coefficients, thereby improving the adaptability of the Gaussian process under different wear levels and different surface conditions.

[0017] Furthermore, the Gaussian process training using the time series of the output real-time sensor data includes: extracting the output real-time sensor data for a predetermined number of consecutive sampling periods prior to the current moment as a time series segment; inputting the time series segment into a long short-term memory network to obtain a hidden layer state vector, and performing dimensionality reduction processing on the hidden layer state vector to obtain a scalar process state index; using the input CNC machining parameters and the geometric latent variables of the first latent space as the independent variables of the Gaussian process, and using the scalar process state index as the dependent variable of the Gaussian process; using the independent variables and the dependent variables to update the hyperparameters of the Gaussian process through maximum likelihood estimation to optimize the kernel parameters of the covariance function and complete model training.

[0018] Furthermore, constructing the parent latent variable with a prior Gaussian mixture distribution includes: setting discrete gradients for cutting depth and feed rate, running the cutting mechanics model under 32 gradient combinations to generate a multidimensional simulation dataset; using the K-Means algorithm to calculate the distance between samples in the multidimensional simulation dataset, and clustering the samples into ten preset classification clusters; calculating the core average values ​​of the sample features within each of the ten classification clusters, and mapping the obtained core average values ​​to the dimensions of the parent latent variable through a transition transformation matrix, which are then used as the initial means of the corresponding components of the Gaussian mixture distribution.

[0019] Furthermore, the transition matrix is ​​obtained through pre-training of an autoencoder, which includes an encoder and a decoder. The encoder is used to compress the 15-dimensional feature vector in the multidimensional simulation dataset into a 5-dimensional latent variable vector, and the decoder is used to reconstruct the 15-dimensional feature vector based on the 5-dimensional latent variable vector. During training, the 15-dimensional feature vector is input into the encoder to obtain the 5-dimensional latent variable vector, and then the 5-dimensional latent variable vector is input into the decoder to obtain the reconstructed feature vector. The mean square error between the reconstructed feature vector and the original 15-dimensional feature vector is used as the loss function for training. After training, the weight parameters used to map from the 15-dimensional feature space to the 5-dimensional latent variable space are extracted from the encoder to form the transition matrix.

[0020] Further, determining the optimization objective function includes: integrating the product of all probability densities of the posterior prediction distribution of processing quality and the corresponding processing quality data to obtain the expected value; finding the boundary position in the posterior prediction distribution of processing quality where the cumulative probability reaches a preset percentage, and extracting the maximum processing quality value corresponding to the boundary position as the risk quantile; extracting the geometric latent variable corresponding to the area to be processed in the first latent space, calculating the L1 norm of the geometric latent variable as a local geometric complexity index, inputting the local geometric complexity index into a nonlinear activation function to map it to a set interval, taking the mapped value as the weight coefficient of the risk quantile, and subtracting the weight coefficient of the risk quantile from 1 to obtain the weight coefficient of the expected value; assigning corresponding weight coefficients to the risk quantile and the expected value respectively, calculating their product and summing them to generate the optimization objective function to be minimized.

[0021] The beneficial effects of this invention are as follows: By using manifold learning, the geometric feature vectors of the components are mapped to the first latent space to preserve the geodesic constraints of the local surfaces, and the CNC machining parameters and geometric latent variables are mapped to the second latent space through a Gaussian process. Combined with the kernel length scale determined by the linkage between the tool wear state evaluation value and the geometric latent variables, the nonlinear response of the cutting process and the tool state are uniformly characterized, avoiding the inaccuracy of prediction under different working conditions by the fixed model; the cutting mechanics laws are integrated into the initialization of the parent latent variables through the hybrid Gaussian prior constructed by clustering and autoencoder, improving the stability of quality prediction; in the parameter optimization stage, the risk quantile and expected value weighted sum of the posterior prediction distribution of machining quality are used as the optimization objective, and the weights are adaptively allocated according to the geometric complexity, thereby taking into account the average accuracy in the smooth region and strengthening the prevention of extreme deviations in the abrupt region, reducing the fluctuation of machining deviation, and improving the consistency of multi-axis CNC machining quality of automotive precision components and the reliability of parameter optimization. Attached Figure Description

[0022] Figure 1This is a schematic flowchart illustrating the adaptive optimization method for multi-axis CNC machining parameters of automotive precision components according to an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram illustrating the timing curve of the absolute deviation of the servo current in an embodiment of the present invention; Figure 3 This is a schematic illustration of a K-Means clustering scatter plot of multidimensional mechanical simulation data in an embodiment of the present invention; Figure 4 This is a schematic bar chart showing the comparison of the ablation test results of processing deviations in an embodiment of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] Example of an adaptive optimization method for multi-axis CNC machining parameters of automotive precision components: like Figure 1 As shown, the adaptive optimization method for multi-axis CNC machining parameters of automotive precision components of the present invention includes: S1, acquire historical machining data, which includes input CNC machining parameters and input component geometric feature vectors, as well as time series data including output machining quality data and output real-time sensor data.

[0026] The machining data of the CNC machine tool is collected via the MTConnect protocol through an industrial IoT gateway and stored in the InfluxDB time-series database. Historical machining data is read, and the input CNC machining parameters include feed rate and depth of cut. Surface geometry analysis is performed on the STEP or STL format 3D model file of the component to be machined, extracting the normal vector angles and principal curvatures of local surfaces, and concatenating them according to a preset feature order to form a multi-dimensional and one-dimensional feature array. Geometric features of different dimensions, such as normal vector angles and principal curvatures, are normalized to obtain dimensionless component geometric feature vectors. The output machining quality data is obtained by using a coordinate measuring machine to obtain the average absolute deviation between the actual contour points and the theoretical positioning coordinates, and imported into a CSV file. The time series of output real-time sensor data is acquired by a piezoelectric triaxial accelerometer and current sensor mounted on the machine tool spindle, and noise reduction and Z-normalization preprocessing are performed using a Butterworth low-pass filter function to construct a structured time-series floating-point matrix.

[0027] In an optional embodiment, historical machining data is acquired, which includes input CNC machining parameters and input component geometric feature vectors, as well as a time series containing output machining quality data and output real-time sensor data, including: The three-phase current signal of the servo drive of the CNC machine tool is collected, and the absolute deviation of the current in each sampling period is calculated to form the time series of the output real-time sensor data. Extract the angle between the normal vectors and the principal curvature of the surface of the component's 3D model to generate the input component's geometric feature vector; Obtain the numerical combination of feed rate and depth of cut as input CNC machining parameters; The machined area is scanned using a coordinate measuring machine, and the average absolute deviation between the actual contour points and the theoretical positioning coordinates is calculated as the output machining quality data.

[0028] Data acquisition cards are used during the processing. Sampling rate acquisition of CNC machine tools , , Three-axis servo drive , , The three-phase current signal is used to calculate the absolute difference between the measured current value and the reference value under rated no-load conditions in real time within each sampling period. For example, the difference is extracted as follows: , The data points are then connected in chronological order to obtain a time series of tens of thousands of points, representing the microscopic fluctuations of the cutting load. The servo current absolute deviation time series curve is shown below. Figure 2 As shown.

[0029] The STEP or STL format 3D model file of the component to be processed is parsed to obtain the geometric information of the surface patches or triangular patches of the component. The normal vector angles between adjacent patches are traversed and extracted, ranging from 0 to 180 degrees (e.g., 45 degrees in the example), as well as the maximum and minimum principal curvatures of each vertex (e.g., 0.02 mm and -0.01 mm). These normal vector angles and principal curvatures are normalized and then concatenated to generate geometric feature vectors, used to describe the local morphology of the workpiece surface. For CNC machining parameters, a continuous combination of feed rate and depth of cut preset in the machining program is read from the machine tool controller, such as a feed rate of 800 mm / min and a depth of cut of 1.5 mm. This two-dimensional array is used as the core parameter input to the model. After processing, a coordinate measuring machine is used in a constant temperature measuring room of 20 degrees Celsius to perform contact scanning on the processed curved surface area at a point cloud spacing of 2 mm. The average absolute deviation between the measured point cloud contour and the normal distance of the measured point theoretical positioning coordinate plane is calculated based on the least squares method. For example, the average absolute deviation is 0.035 mm. This contour error index is used as the processing quality data output by the model training.

[0030] S2. Construct a hierarchical latent variable model to establish a mapping relationship from input to output. The process of constructing the hierarchical latent variable model includes mapping the input component geometric feature vectors to the first latent space through a manifold learning algorithm, and combining the input CNC machining parameters with the geometric latent variables in the first latent space, which are then mapped to the second latent space through a Gaussian process. The Gaussian process is trained using the time series of the output real-time sensor data and determines the tool wear state assessment value based on tool usage information. The kernel length scale of the covariance function of the Gaussian process is determined by the correlation between the tool wear state assessment value and the geometric latent variables. Parent latent variables with a prior Gaussian mixture distribution are constructed in the first and second latent spaces, and a mapping between the parent latent variables and the output machining quality data is established.

[0031] In constructing the hierarchical latent variable model, the Isomap manifold learning algorithm is used to reduce the dimensionality of the high-dimensional component geometric feature vectors, mapping them to a low-dimensional first latent space to obtain geometric latent variables that characterize the geometric constraints of the local surface, while preserving the intrinsic geometric geodesic constraints of the surface. The CNC machining parameters and these geometric latent variables are tensor-concatenated as joint independent variable inputs to the Gaussian process. A Gaussian process regression model is constructed and mapped to a second latent space to represent the nonlinear characteristics of the cutting process. The scalar process state index, obtained by dimensionality reduction of the real-time sensor time series extracted through a long short-term memory network, is used as the training dependent variable for the Gaussian process.

[0032] When setting the Matern covariance function for the Gaussian process, the current tool wear state assessment value is calculated using the pre-established Taylor tool life equation. Let the Taylor tool life equation be... ,in The current cutting speed, This represents the theoretical tool life at the cutting speed. and This is a life coefficient pre-calibrated based on tool material, workpiece material, and cutting conditions. The current cutting speed is calculated based on the machine tool spindle speed and tool diameter. , and by Calculate the theoretical tool life; read the current cumulative cutting time of the tool. ,calculate and will The tool wear condition is then limited to the range of 0 to 1 to obtain the current tool wear condition assessment value. The closer the tool wear condition assessment value is to 1, the higher the degree of tool wear.

[0033] The kernel length scale of the covariance function of the Gaussian process is determined by the correlation between the tool wear condition assessment value and the geometric latent variable. Specifically, for the geometric latent variable... Standardize the process; evaluate the tool wear value. and Concatenate into a joint vector And calculate intermediate variables. ,in For the preset weight vector, Preset bias; scale adjustment coefficient ,in and These are the preset lower and upper limits of the adjustment coefficient, respectively, and both are positive numbers; the basic kernel length scale and the scale adjustment coefficient are used. Multiply to obtain the current kernel length scale, and use the current kernel length scale to update the covariance matrix of the Gaussian process.

[0034] The output feature vectors of the first and second latent spaces are concatenated into channel pairs and used as the bottom-level observation variables. A parent latent variable node, with a prior set as a Gaussian Mixture Model (GMM), is trained using the Expectation-Maximization (EM) algorithm or an approximate variational inference process. A multilayer perceptron (MLP) neural network with three fully connected layers is constructed, using the parent latent variable as the input feature and the processing quality data as the output label. End-to-end joint backpropagation training is performed using the Adaptive Moment Estimation (Adam) algorithm and the Mean Squared Error (MSE) loss function until the model's loss value converges on the validation set, thus establishing a nonlinear mapping relationship from multi-source heterogeneous inputs to processing quality outputs.

[0035] In an optional embodiment, mapping the input component geometric feature vector to a first latent space using a manifold learning algorithm includes: A K-nearest neighbor graph is constructed based on the input component geometric feature vectors, and the Euclidean distance between nodes is calculated as the initial edge weights. Dijkstra's algorithm is used to calculate the shortest path between any two nodes in the graph and construct the geodesic distance matrix. A multidimensional scaling transformation is performed on the geodesic distance matrix. The low-dimensional embedding coordinates of each sample are calculated based on the first three eigenvalues ​​and their corresponding eigenvectors. The low-dimensional embedding coordinates are then used as geometric latent variables mapped to the first latent space.

[0036] For the component geometric feature vectors containing the normal angle and principal curvature, the nearest neighbor number K is set to 10. In the feature space, the 10 most similar nearest neighbors of each patch node are found, constructing a locally connected undirected K-nearest neighbor graph. The Euclidean distance between connected node vectors is calculated, such as an example value of 1.25, as the initial edge weight of the graph, preserving the local neighborhood structure of the features. Based on this weighted nearest neighbor graph, Dijkstra's algorithm is used to traverse the graph structure, finding and accumulating the minimum edge weight path sum between each pair of non-adjacent nodes, generating an N×N geodesic distance matrix containing the global manifold structure, such as a geodesic distance of 5.68 between any two endpoints. A double centering operation is performed on the constructed geodesic distance matrix to generate an inner product matrix, and eigenvalue decomposition is performed through multidimensional scaling transformation to expand the manifold into a low-dimensional Euclidean space. After calculating and sorting all eigenvalues ​​of the inner product matrix, the top three largest eigenvalues ​​are extracted, such as 45.2, 12.5, and 8.1, and their corresponding three eigenvectors. The coordinates of each sample in the three-dimensional low-dimensional embedding space are calculated based on the first three eigenvalues ​​and their corresponding eigenvectors, such as the example vectors 0.85, -0.42, and 0.17. This vector is then used as a geometric latent variable mapped to the first latent space, which represents the local surface topological constraints on the machining toolpath while compressing the data dimension.

[0037] In an optional embodiment, the Gaussian process is trained using time series of the output real-time sensor data, including: Extract real-time sensor data from the output of a preset number of consecutive sampling periods prior to the current moment as a time series segment; Time series segments are input into a long short-term memory network to obtain hidden state vectors, and the hidden state vectors are then dimensionality-reduced to obtain scalar process state indices. The input CNC machining parameters and the geometric latent variables of the first latent space are used as the independent variables of the Gaussian process, and the scalar process state index is used as the dependent variable of the Gaussian process. The hyperparameters of the Gaussian process are updated using maximum likelihood estimation with independent and dependent variables to optimize the kernel parameters of the covariance function and complete model training.

[0038] During the time series data processing stage, a sliding window mechanism is used to extract a data segment from the servo current absolute deviation sequence, with the current time as the endpoint and a length set to 20 sampling periods. For example, to obtain data from... to A real-time sensor time series segment consisting of continuously fluctuating values ​​is input into a Long Short-Term Memory (LSTM) network with 64 hidden layer nodes and Tanh activation function. The LTM network consists of an input layer, hidden layers, and an output layer. The hidden layer contains multiple memory units expanded over time steps, and each memory unit contains a forget gate, an input gate, an output gate, and a cell state. The input to the LTM network is a time series tensor consisting of sensor data from the aforementioned twenty consecutive sampling periods. The output of the LTM network is the hidden layer state vector of the last time step, with a dimension equal to the number of hidden layer nodes (64 dimensions).

[0039] Long Short-Term Memory (LSTM) networks perform time-step forward propagation on time-series segments. At each time step, the network calculates the gating values ​​for the forget gate, input gate, and output gate based on the current input data vector and the hidden state vector from the previous time step. The forget gate controls the retention rate of cell states from the previous time step, the input gate controls the proportion of current input information written into cell states, and candidate cell states represent new state information that can be written to the current time step. Subsequently, the network performs element-wise weighted fusion of the cell states from the previous time step and the candidate cell states according to the gating results of the forget gate and the input gate to obtain the cell state of the current time step. Then, it filters the current cell states based on the gating result of the output gate to generate the hidden state vector of the current time step. After 20 consecutive sampling cycles of time-step iterations, the 64-dimensional hidden state vector output from the last time step is extracted as a process representation of the cumulative effects of cutting vibration, load fluctuation, and tool wear. The 64-dimensional hidden state vector is then converted into a scalar process state index through a fully connected dimensionality reduction layer or average pooling.

[0040] A Gaussian process regression model is constructed. Two-dimensional CNC machining parameters (feed rate and depth of cut, normalized to the 0-1 interval) are concatenated with geometric latent variables in the first latent space and used as the input independent variable matrix. The aforementioned scalar process state indicators are output as the dependent variable, resulting in training data pairs. The L-BFGS-B quasi-Newton optimization algorithm is employed, using the maximization of the logarithmic marginal likelihood function as the criterion, to perform maximum likelihood estimation with a maximum of 500 iterations. During this iteration process, the fundamental kernel hyperparameters in the squared exponential covariance function or Matern function are updated. For example, the signal variance parameter is optimized to converge to 1.2, the noise variance parameter to 0.01, and the fundamental feature length scale parameter to 0.5. When the model calculates the covariance of the current sample, a scale adjustment coefficient is calculated based on the tool wear state assessment value and the geometric latent variables. This scale adjustment coefficient is then applied to the fundamental feature length scale to obtain the effective kernel length scale corresponding to the current sample. This establishes the mapping relationship between CNC machining parameters and geometric latent variables to process state features, completing the training of the Gaussian process model.

[0041] After training, when new CNC machining parameters and corresponding first latent space geometric latent variables are input, the Gaussian process model outputs the corresponding process state latent variables. The process state latent variables constitute the representation of the sample in the second latent space, which is used to characterize the dynamic response of the cutting process under the corresponding machining conditions.

[0042] In an optional embodiment, parent latent variables with a prior Gaussian mixture distribution are constructed in the first and second latent spaces, including: Discrete gradients for depth of cut and feed rate are set, and a cutting mechanics model is run under 32 gradient combinations to generate a multidimensional simulation dataset; The K-Means algorithm is used to calculate the distance between samples in the simulation dataset and cluster the samples into ten preset classification clusters. Calculate the core mean values ​​of the sample features within each of the ten classification clusters. Then, map each core mean value to the parent latent variable dimension through a transition matrix and use them as the initial mean values ​​of the corresponding components of the mixture Gaussian distribution.

[0043] The cutting process parameter domain was discretized, with four cutting depth values ​​set at 0.5 mm, 1.0 mm, 1.5 mm, and 2.0 mm; and eight feed rates set ranging from 200 mm / min to 1600 mm / min, with a step size of 200 mm / min. Orthogonal combinations of cutting depth and feed rate were performed to obtain 32 sets of parameter gradient conditions. Each set of parameter conditions was input into finite element cutting simulation software to establish a three-dimensional thermo-mechanical coupled cutting finite element model corresponding to the workpiece material, tool material, tool geometry, and boundary conditions. The cutting process simulation was performed under the same simulation time, mesh generation accuracy, and contact friction conditions. After the simulation stabilized, 15 simulation result features, including triaxial cutting force, maximum cutting temperature, shear band stress, tool contact zone pressure, and equivalent plastic strain, were extracted for each parameter condition and arranged into a 15-dimensional feature vector according to a preset order. The above simulation and feature extraction were performed on each of the 32 sets of conditions to generate a multi-dimensional mechanical simulation dataset containing 32 samples, each corresponding to a 15-dimensional feature vector. A three-dimensional thermo-mechanical coupled cutting finite element model refers to a simulation model that simultaneously simulates material deformation, cutting force transmission, and temperature field evolution during the cutting process in a three-dimensional finite element space, and considers the mutual influence between mechanical deformation and cutting heat.

[0044] The multidimensional mechanical simulation dataset was input into the K-Means clustering algorithm. The number of clusters was set to 10, and Euclidean distance was used as the sample similarity metric. With a maximum iteration count of 1000 and a cluster center drift tolerance of 0.001, the samples in the multidimensional simulation dataset were divided into 10 non-overlapping processing mode clusters. After clustering, the arithmetic mean of the 15-dimensional feature vectors within each of the 10 clusters was calculated, yielding 10 15-dimensional core mean vectors representing the centrality of each specific working condition cluster. The K-Means clustering scatter plot of the multidimensional mechanical simulation data after principal component analysis (PCA) dimensionality reduction is shown below. Figure 3 As shown.

[0045] A pre-trained transition matrix obtained through an autoencoder is used to map 15-dimensional mechanical simulation features to a 5-dimensional parent latent variable space. The autoencoder consists of an encoder and a decoder. The encoder compresses the 15-dimensional feature vectors in the multi-dimensional mechanical simulation dataset into 5-dimensional latent variable vectors, and the decoder reconstructs the 15-dimensional feature vectors from the 5-dimensional latent variable vectors. During training, the 15-dimensional feature vectors are input into the encoder to obtain the 5-dimensional latent variable vectors, and then the 5-dimensional latent variable vectors are input into the decoder to obtain the reconstructed feature vectors. The mean squared error between the reconstructed feature vectors and the original 15-dimensional feature vectors is used as the loss function for training. After training, the weight parameters used in the encoder to map from the 15-dimensional feature space to the 5-dimensional latent variable space are extracted to form the transition matrix.

[0046] Each 15-dimensional core mean vector is multiplied by the transition matrix to reduce its dimensionality and map it to a parent latent variable space of dimension 5. The resulting 10 5-dimensional mapping vectors, such as example vectors 1.2, -0.5, 0.8, 0.3, and -1.1, are then assigned the initial mean parameters of the 10 independent Gaussian components in the target Gaussian mixture model.

[0047] S3. Input the geometric feature vector of the area to be processed into the trained hierarchical latent variable model, and optimize the CNC machining parameters through backpropagation. The optimization objective function is the weighted sum of the risk quantile and the expected value of the posterior prediction distribution of the machining quality, and the weights of the weighted sum are adjusted according to the representation of the geometric feature vector in the first latent space.

[0048] In the CNC machining parameter optimization execution stage, the geometric feature vector of the new region of the automotive component to be processed is extracted and input into the hierarchical latent variable model. The input CNC machining parameters are initialized to the basic values ​​recommended in the process manual, and these basic values ​​are set as the learnable tensors for which gradients need to be calculated. When constructing the optimization objective function, forward propagation is performed based on the Gaussian process posterior distribution, the mixture of Gaussian priors of the parent latent variables, and the quality mapping network to obtain the probabilistic posterior prediction distribution of the machining quality. The risk quantile of the posterior prediction distribution at a preset confidence level is calculated using the inverse cumulative distribution function, and the mean of this distribution is calculated as the expected value of the conventional machining quality. A single-layer perceptron or nonlinear activation function is used to perform feature mapping on the geometric latent variables of the geometric feature vector in the first latent space, and the output is a risk weight coefficient mapped to a value between 0.1 and 0.9. The risk weight coefficient is used as the weight of the risk quantile term, and the complementary value obtained by subtracting the risk weight coefficient from 1 is used as the weight of the expected value term to adjust the proportional allocation in the objective weighted sum formula.

[0049] Keeping all neural network and kernel parameters of the hierarchical latent variable model constant, and guided by minimizing the weighted sum objective function, the LBFGS nonlinear optimization algorithm is employed to perform backpropagation differentiation and step-size iterative updates on the CNC machining parameter input tensor. When the gradient of the objective function... When the norm is less than the set small convergence threshold or the set maximum number of iterations is reached, the optimization calculation stops, the tensor data type is converted into a floating-point number, the optimized combination of CNC machining parameters is output, and the best combination of CNC machining parameters is sent to the main control program of the Siemens 840D CNC system for machining through the OPCUA communication protocol.

[0050] In an optional embodiment, the objective function is optimized as a weighted sum of the risk quantiles and expected values ​​of the posterior prediction distribution of processing quality, including: The expected value is obtained by integrating the product of the entire probability density of the posterior prediction distribution of the output processing quality data and the corresponding processing quality data. In the posterior prediction distribution, find the boundary position where the cumulative probability reaches a preset percentage, and extract the maximum processing quality value corresponding to the position as the risk quantile. After assigning appropriate weight coefficients to the risk quantile and the expected value, the product of the two is calculated and summed to generate the optimization objective function that needs to be minimized.

[0051] For the posterior Gaussian prediction distribution of the machining quality, i.e. the average absolute deviation, output by the hierarchical latent variable model, the probability density function value of the prediction distribution is multiplied with the corresponding continuous variable of machining deviation within a set absolute deviation range, such as 0 to 0.5 mm, and the definite integral is calculated to obtain the expected value of the prediction distribution. For example, the expected machining deviation value is calculated to be 0.025 mm, which represents the average machining accuracy under normal working conditions.

[0052] To prevent extreme processing degradation, an inverse cumulative distribution function (IFD) calculation mechanism is introduced. The cumulative probability area is calculated from left to right on the posterior prediction curve of processing quality, locating the unilateral boundary where the cumulative probability reaches 99%, and extracting the maximum processing quality limit corresponding to this boundary, such as 0.045 mm as the processing deviation risk quantile. After extracting the expected value of 0.025 mm representing the average state and the risk quantile of 0.045 mm representing the extreme state, specific non-negative weight coefficients are assigned to these two indicators, for example, a bias weight of 0.7 for the risk quantile and a base weight of 0.3 for the expected value. A linear combination is then performed, calculating the sum of the products of the two indicators and their corresponding weights, i.e., 0.7 x 0.045 + 0.3 x 0.025 equals 0.039 mm. This weighted result, which integrates both stability and average accuracy indicators, is set as the optimization objective function to be minimized during the reverse optimization process. Based on this, the feed rate and depth of cut combination that minimizes this weighted objective value is searched.

[0053] In an optional embodiment, the weights of the weighted sum are adjusted based on the representation of the geometric feature vector in the first latent space, including: Extract the geometric latent variables corresponding to the region to be processed in the first latent space, and calculate the L1 norm of the geometric latent variables as a local geometric complexity index. The local geometric complexity index is input into a nonlinear activation function to map the value to a set interval, and the mapped value is taken as the weight coefficient of the risk quantile. The weighting coefficient of the expected value is obtained by subtracting the weighting coefficient of the risk quantile from 1, thus realizing the allocation of weights.

[0054] For the surface mesh region to be machined before cutting, the corresponding 3D geometric latent variable vector is extracted from the first latent space of the hierarchical latent variable model. For example, if the vector extraction results are 1.25, -0.80, and 0.45, the absolute values ​​of each element of this vector are calculated and then scalarly summed. That is, the absolute value of 1.25 plus the absolute value of -0.80 plus the absolute value of 0.45 equals 2.50. The geometric latent variable is then calculated. The norm is defined and set as a local geometric complexity index to assess the degree of curvature abrupt change in the current region.

[0055] The obtained The norm value is input into a nonlinear activation function with scaling and translation properties, such as a translational sigmoid function, mapping the local geometric complexity value to a defined closed interval of 0.1 to 0.9. Assuming the calculated nonlinear mapping value is 0.76, this 0.76 is set as a high-priority weight coefficient for the risk quantile in the objective function, causing the optimizer to tend to avoid processing errors with a greater weight when facing geometric surfaces such as steep transitions. Utilizing the principle of complementary probability weights, the risk quantile weight coefficient of 0.76 is subtracted from a constant 1 to obtain a remaining value of 0.24, which is used as the associated weight coefficient for the expected value. This linkage mechanism based on the geometric manifold characteristics of the first latent space allows the weight parameters to be allocated between 0.1 and 0.9, helping the algorithm to increase its focus on optimizing the expected average accuracy in smooth regions and increasing its focus on preventing maximum limit risks in abrupt regions.

[0056] The experiment was conducted under identical CNC machining environment and measurement standards, machining one hundred standard titanium alloy curved surface test pieces. The machine tool spindle speed was uniformly set to 3,000 revolutions per minute, and carbide ball end mills of the same material and specifications were used. All models were trained and parameter optimized on a workstation with the same graphics processor configuration. After machining, the machined surface area was scanned at a point cloud spacing of two millimeters, and the average absolute deviation and maximum extreme deviation of the machined surface were statistically analyzed as the core evaluation indicators.

[0057] The baseline model using only static weights and traditional feature inputs yielded an average absolute deviation of 0.052 mm and a maximum extreme deviation of 0.086 mm. The ablation model removing adaptive weighting and using a fixed weight of 0.5 yielded an average absolute deviation of 0.041 mm and a maximum extreme deviation of 0.065 mm. The ablation model removing Gaussian priors and using random initialization yielded an average absolute deviation of 0.044 mm and a maximum extreme deviation of 0.071 mm. The model fully applying all functional modules of this invention yielded an average absolute deviation of 0.031 mm and a maximum extreme deviation of 0.043 mm. Comparison of processing deviation ablation experimental results is provided. Figure 4 As shown.

[0058] The complete solution reduces the average absolute deviation by approximately [percentage missing] compared to the baseline model. And reduce the maximum extreme deviation by approximately The above results show that the mixed Gaussian distribution prior constructed using clustering and autoencoders can incorporate cutting mechanics laws to a certain extent, improving the stability of model initialization and quality prediction. Compared with the static weight variant, the geometric complexity linkage mechanism based on the first latent space extraction of the manifold helps to detect local topological features of the workpiece and assigns higher optimization weights to the risk quantile in steep abrupt change regions, thereby reducing the probability of extreme out-of-tolerance and improving the consistency of cutting quality.

Claims

1. An adaptive optimization method for multi-axis CNC machining parameters of automotive precision components, characterized in that, Includes the following steps: Acquire historical processing data, which includes input CNC machining parameters and input component geometric feature vectors, as well as a time series containing output processing quality data and output real-time sensor data; A hierarchical latent variable model is constructed to establish a mapping relationship from input to output. The process of constructing the hierarchical latent variable model includes mapping the geometric feature vector of the input components to a first latent space through a manifold learning algorithm, combining the input CNC machining parameters with the geometric latent variables of the first latent space, and mapping them to a second latent space through a Gaussian process. The Gaussian process is trained using the time series of the output real-time sensor data and determines the tool wear state assessment value based on tool usage information. The kernel length scale of the covariance function of the Gaussian process is determined by associating the tool wear state assessment value with the geometric latent variables. Parent latent variables with a prior Gaussian mixture distribution are constructed in the first latent space and the second latent space, and a mapping is established between the parent latent variables and the output machining quality data. For the geometric latent variables Standardization processing is required; The tool wear condition assessment value and Concatenate into a joint vector Calculate intermediate variables ,in For the preset weight vector, This is the preset offset; T is the theoretical tool life at the current cutting speed. according to Calculate the scaling adjustment coefficient ,in and These are the preset lower and upper limits of the adjustment coefficient, both of which are positive numbers; The basic kernel length scale is compared with the scale adjustment coefficient. Multiply to obtain the current kernel length scale, and use the current kernel length scale to update the covariance matrix of the Gaussian process; The geometric feature vector of the area to be processed is input into the hierarchical latent variable model that has been trained, and the CNC machining parameters are optimized through backpropagation. The optimization objective function is the weighted sum of the risk quantile and the expected value of the posterior prediction distribution of the machining quality, and the weights of the weighted sum are adjusted according to the representation of the geometric feature vector in the first latent space. The parent latent variables that are constructed prior to a Gaussian mixture distribution include: Discrete gradients for depth of cut and feed rate are set, and a cutting mechanics model is run under 32 gradient combinations to generate a multidimensional simulation dataset; The K-Means algorithm is used to calculate the distance between samples in the multidimensional simulation dataset, and the samples are clustered into ten preset classification clusters; Calculate the core average values ​​of the sample features within each of the ten classification clusters. After mapping each core average value to the dimension of the parent latent variable through a transition transformation matrix, use them as the initial average values ​​of the corresponding components of the Gaussian mixture distribution. The determination of the optimization objective function includes: The expected value is obtained by integrating the product of all probability densities of the posterior prediction distribution of the processing quality with the corresponding processing quality data; the boundary position where the cumulative probability reaches a preset percentage is found in the posterior prediction distribution of the processing quality, and the maximum processing quality value corresponding to the boundary position is extracted as the risk quantile. Extract the geometric latent variables corresponding to the region to be processed in the first latent space, calculate the L1 norm of the geometric latent variables as a local geometric complexity index, input the local geometric complexity index into a nonlinear activation function to map to a set interval, take the mapped value as the weight coefficient of the risk quantile, and subtract the weight coefficient of the risk quantile from 1 to obtain the weight coefficient of the expected value; assign corresponding weight coefficients to the risk quantile and the expected value respectively, calculate the product of the two and sum them to generate the optimization objective function to be minimized.

2. The adaptive optimization method for multi-axis CNC machining parameters of automotive precision components according to claim 1, characterized in that, When acquiring the historical processing data, the following are included: The three-phase current signal of the servo drive of the CNC machine tool is collected, and the absolute deviation of the current in each sampling period is calculated to form the time series of the output real-time sensor data. Extract the angle between the normal vectors and the principal curvature of the surface of the component's 3D model to generate the input component geometric feature vector; The numerical combination of feed rate and depth of cut is obtained as the input CNC machining parameters; the machined area is scanned using a coordinate measuring machine, and the average absolute deviation between the actual contour points and the theoretical positioning coordinates is calculated as the output machining quality data.

3. The adaptive optimization method for multi-axis CNC machining parameters of automotive precision components according to claim 1, characterized in that, The manifold learning algorithm is the Isomap algorithm, which is used to reduce the dimensionality of the input component geometric feature vector and map it to the first latent space to obtain geometric latent variables that characterize the local surface geometric constraints. The input CNC machining parameters and the geometric latent variables are tensor-concatenated and used as the joint independent variable input of the Gaussian process and mapped to the second latent space; the time series of the output real-time sensor data is extracted by a long short-term memory network and then dimensionality-reduced to obtain a scalar process state index, which is used as the training dependent variable of the Gaussian process.

4. The adaptive optimization method for multi-axis CNC machining parameters of automotive precision components according to claim 3, characterized in that, Mapping to the first latent space using the Isomap algorithm includes: A K-nearest neighbor graph is constructed based on the input component geometric feature vectors, and the Euclidean distance between nodes is calculated as the initial edge weights. Dijkstra's algorithm is used to calculate the shortest path between any two nodes in the graph and construct the geodesic distance matrix. A multidimensional scaling transformation is performed on the geodesic distance matrix, and the low-dimensional embedding coordinates of each sample are calculated based on the first three eigenvalues ​​and their corresponding eigenvectors. The low-dimensional embedding coordinates are then used as geometric latent variables mapped to the first latent space.

5. The adaptive optimization method for multi-axis CNC machining parameters of automotive precision components according to claim 1, characterized in that, Determining the tool wear condition assessment value includes: Using the pre-established Taylor tool life equation Calculate the current tool wear condition assessment value, where The current cutting speed, and This is a life coefficient pre-calibrated based on the tool material, workpiece material, and cutting conditions; Calculate the current cutting speed based on the machine tool spindle speed and tool diameter. , and by Calculate the theoretical tool life; read the current cumulative cutting time of the tool. ,calculate and will Limit the amplitude to the range of 0 to 1 to obtain the current tool wear status assessment value.

6. The adaptive optimization method for multi-axis CNC machining parameters of automotive precision components according to claim 1, characterized in that, The Gaussian process is trained using the time series of the output real-time sensor data, including: Extract the real-time sensor data output from the current moment for a predetermined number of consecutive sampling periods as a time series segment; The time series segment is input into a long short-term memory network to obtain a hidden state vector, and the hidden state vector is dimensionality reduced to obtain a scalar process state index. The input CNC machining parameters and the geometric latent variables of the first latent space are used as the independent variables of the Gaussian process, and the scalar process state index is used as the dependent variable of the Gaussian process. The hyperparameters of the Gaussian process are updated using the independent variables and the dependent variables through maximum likelihood estimation to optimize the kernel parameters of the covariance function and complete the model training.

7. The adaptive optimization method for multi-axis CNC machining parameters of automotive precision components according to claim 6, characterized in that, The transition matrix is ​​obtained through pre-training of an autoencoder, which includes an encoder and a decoder. The encoder is used to compress the 15-dimensional feature vector in the multidimensional simulation dataset into a five-dimensional latent variable vector, and the decoder is used to reconstruct the 15-dimensional feature vector based on the five-dimensional latent variable vector. During training, the 15-dimensional feature vector is input into the encoder to obtain a five-dimensional latent variable vector, and then the five-dimensional latent variable vector is input into the decoder to obtain a reconstructed feature vector. The mean square error between the reconstructed feature vector and the original 15-dimensional feature vector is used as the loss function for training. After training, the weight parameters used to map from the 15-dimensional feature space to the five-dimensional latent variable space in the encoder are extracted to form the transition matrix.

Images