Machine learning-based automobile part process parameter real-time optimization method and system

Abstract

The invention relates to the technical field of automobile part processing, and discloses an automobile part process parameter real-time optimization method and system based on machine learning, and the method comprises the steps: collecting the temperature gradient, pressure distribution, cutting speed and other multi-source process parameter data, extracting a process feature sequence through a self-attention mechanism and a time convolution network, and carrying out the real-time optimization of the process feature sequence; calculating a process fluctuation coefficient; and generating a process correlation weight through covariance matrix characteristic decomposition and Sigmoid function transformation, dynamically adjusting a reference parameter to generate an optimized parameter, and regulating and controlling equipment operation. The system comprises a multi-source data acquisition module, a process feature extraction module, a fluctuation coefficient calculation module and the like. And a multi-dimensional process parameter space is also constructed to monitor an abnormal state, and a joint optimization model is established to realize collaborative optimization of process and equipment control parameters. The real-time performance and accuracy of technological parameter optimization are improved, machining error accumulation is effectively restrained, the method is suitable for intelligent machining of automobile parts, and the machining quality and efficiency are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of automobile parts processing, and in particular to a method and system for real-time optimization of automobile parts process parameters based on machine learning. Background Art

[0002] In the automotive manufacturing sector, the quality of component processing directly affects the performance and safety of the entire vehicle, and the precise control of process parameters is the core link to ensure processing quality. Traditional optimization of process parameters for automotive component processing mainly relies on manual experience or offline simulation, which has significant limitations. Manual experience optimization is greatly affected by the operator's professional level and subjective judgment, and it is difficult to cope with complex processing scenarios under the coupling of multiple processes and multiple parameters. In addition, the optimization cycle is long and the efficiency is low, which cannot meet real-time requirements. Although offline simulation can simulate the processing process by establishing a mathematical model, the model construction relies on a large number of assumptions and simplifications, which are different from the actual processing environment. In particular, when faced with dynamic changes in parameters such as temperature gradients, pressure distribution, and cutting speed during the processing process, the accuracy and reliability of the simulation results are greatly reduced.

[0003] As the automotive industry develops toward intelligent and high-precision manufacturing, machining processes are characterized by the synergy of multiple parameters and the close integration of multiple processes, making traditional optimization methods difficult to adapt. For example, in the multi-process machining of complex parts, fluctuations in the process parameters of the previous process are transmitted to the subsequent processes through factors such as equipment load and workpiece material properties, forming a parameter coupling effect. If the interrelated influence of the process parameters of each process cannot be monitored and adjusted in real time, it is easy to lead to the accumulation of machining errors, resulting in problems such as excessive dimensional accuracy of parts and deterioration of surface quality, increasing scrap rate and production costs.

[0004] The development of machine learning technology has provided a new approach to addressing these issues. The self-attention mechanism can effectively capture long-range dependencies in sequential data and has advantages in processing global feature correlations in multi-source heterogeneous process parameter data. Temporal convolutional networks can extract multi-scale temporal features through dilated convolution kernels, making them suitable for analyzing the dynamic changes in parameters during processing. Data dimensionality reduction techniques such as principal component analysis can extract key features from high-dimensional process parameter data, simplifying computational complexity. However, research on the application of machine learning technology to the real-time optimization of automotive component process parameters is still in the exploratory stage. Further research and practice are needed to organically combine multiple machine learning methods to achieve dynamic correlation analysis of process parameters, real-time optimization, and precise control of processing equipment.

[0005] Furthermore, existing automotive parts processing systems lack real-time monitoring and intelligent response mechanisms for abnormal process parameters. When abnormalities such as excessive density of extreme parameter values ​​or sudden variance changes occur during processing, timely warnings and process parameter adjustments are impossible, potentially leading to equipment failure or quality issues. Furthermore, the coordinated optimization of process parameters and equipment control parameters is insufficient, making it difficult to optimize energy consumption and safely control equipment load while ensuring processing quality. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for real-time optimization of automotive parts process parameters based on machine learning to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a real-time optimization method for automobile parts process parameters based on machine learning, the method comprising: Collecting multi-source process parameter data during the machining of automotive parts, wherein the process parameter data includes temperature gradient, pressure distribution and cutting speed; The self-attention mechanism is used to extract process features from process parameter data to obtain a process feature sequence, and the process fluctuation coefficient is calculated based on the variance change of the process feature sequence; The covariance matrix of the process fluctuation coefficients of different processes is subjected to eigendecomposition to extract the principal component vector, which is then transformed nonlinearly using the Sigmoid function to obtain the process correlation weight of each process. The baseline parameters of the process feature sequence are dynamically adjusted according to the process association weights to generate optimization parameters, and the operating parameters of the processing equipment are regulated based on the optimization parameters.

[0008] Preferably, the process feature extraction includes: Intercept process parameter data at a preset sampling frequency, construct process tensors and perform standardization; A multi-head self-attention layer is used to perform global feature association on the process tensor and output the weighted sum of the attention weight and the feature vector; The attention results are input into the temporal convolutional network, multi-scale temporal features are extracted by dilating the convolution kernel, and the feature channels are screened through the gating mechanism to generate a process feature sequence.

[0009] Preferably, the process fluctuation coefficient calculation includes: Select the process feature sequence of any process as the benchmark parameter and calculate the variance of its m adjacent process feature vectors; The ratio of the mean of the variance to the Manhattan distance of the benchmark parameter is taken as the intra-process fluctuation; Calculate the covariance between the benchmark parameter and the associated process characteristics, and take its absolute value as the inter-process fluctuation; The harmonic mean of the intra-process fluctuation and the inter-process fluctuation is taken as the process fluctuation coefficient.

[0010] Preferably, the principal component vector extraction includes: Construct the covariance matrix of process fluctuation coefficients of different processes and perform eigenvalue decomposition to obtain orthogonal basis vectors; Select the eigenvectors whose eigenvalue variance contribution rate exceeds the set threshold to form the principal component subspace; The covariance matrix is ​​mapped to the principal component subspace to obtain the principal component vector after dimensionality reduction.

[0011] Preferably, the process association weight calculation includes: Normalize the principal component vector and calculate the Pearson correlation coefficient between it and the preset reference vector; The correlation coefficient is input into the deep neural network and the initial weight is generated after transformation by the hidden layer and activation function; The initial weights are smoothed and filtered through a time sliding window to output the process-related weights of each process.

[0012] Preferably, the optimization parameter generation includes: Perform matrix dot multiplication of the process-related weight and the benchmark parameter to obtain the weight adjustment parameter; Calculate the deviation between the weight adjustment parameter and the benchmark parameter, and dynamically correct the deviation through the particle filter; The corrected deviation is added to the baseline parameters to generate the optimized parameters.

[0013] Preferably, the method further comprises: Construct a multidimensional process parameter space based on the optimized parameters, and extract the extreme points and variance mutation intervals of the parameter space; When the extreme point density exceeds the set threshold or the variance mutation interval span is greater than the limit value, it is determined to be a process abnormal state and multi-level control instructions are generated.

[0014] Preferably, the multidimensional process parameter space construction includes: Mapping the optimization parameters to a multi-dimensional coordinate system according to the processing sequence to generate a parameter distribution point set; The kernel density estimation method is used to spatially reconstruct the parameter distribution point set, and the variance gradient and skewness coefficient of the reconstructed space are calculated; The noise of gradient distribution is suppressed by Gaussian filtering algorithm to eliminate random interference.

[0015] Preferably, the method further comprises: Establish a joint optimization model of process parameter data and equipment control parameters, and use sequential quadratic programming to solve the optimal control strategy; The optimal control strategy is coupled with the optimization parameters in real time to generate a process plan with the best processing quality; The joint optimization model solution includes: The objective function is defined as the sum of the weighted absolute values ​​of process parameter deviation and control energy consumption, and the constraint condition is the equipment load safety range; Perform KKT conditional transformation on the objective function and decompose it into control parameter sub-problem and process parameter sub-problem; The two sub-problems are solved alternately and iteratively until convergence, and the optimal control strategy that satisfies the constraints is output.

[0016] Preferably, the present invention further includes a real-time optimization system for automotive parts process parameters based on machine learning, the system comprising: Multi-source data acquisition module, used to obtain temperature gradient, pressure distribution and cutting speed data during the processing in real time, and build a multi-dimensional process parameter time series matrix; The process feature extraction module uses a self-attention mechanism to perform global feature association on the multi-dimensional process parameter time series matrix, captures multi-scale process features through a temporal convolutional network, and outputs a process feature sequence; The process fluctuation coefficient calculation module generates intra-process fluctuation and inter-process fluctuation according to the variance change of the process feature sequence, and calculates their harmonic mean as the process fluctuation coefficient; The covariance matrix decomposition module constructs the covariance matrix for the process fluctuation coefficients of different processes and extracts the principal component vector through eigenvalue decomposition; The weight mapping module uses the Sigmoid function to perform nonlinear transformation on the principal component vector to generate the process-related weights of each process; The parameter optimization module dynamically adjusts the baseline parameters based on process-related weights, corrects parameter deviations through a particle filter, and generates optimized parameters; The real-time control module converts the optimization parameters into control signals for the processing equipment, which are transmitted to the equipment controller via the industrial bus to complete the operation parameter regulation.

[0017] Compared with the prior art, the present invention has the following beneficial effects: In terms of process feature extraction and fluctuation analysis, the self-attention mechanism is used to perform global feature association on multi-source process parameter data such as temperature gradient, pressure distribution, and cutting speed. This can effectively capture the long-distance dependencies between different parameters, overcoming the shortcomings of traditional methods in complex parameter association analysis. Combined with the dilated convolution kernel and gating mechanism of the temporal convolutional network, multi-scale temporal features can be extracted and key feature channels can be screened to generate a process feature sequence that better reflects the essence of the machining process. When calculating the process fluctuation coefficient based on the process feature sequence, the intra-process fluctuation and the inter-process fluctuation are comprehensively considered, which fully reflects the stability of the parameters within the same process and the mutual influence of the parameters between different processes, providing a more accurate basis for subsequent process parameter optimization.

[0018] In terms of calculating process parameter correlation weights and generating optimized parameters, the covariance matrix of the process fluctuation coefficients of different processes is subjected to eigendecomposition, and the principal component vectors are extracted and nonlinearly transformed using the Sigmoid function. This quantifies the degree of process correlation between each process and generates reasonable process correlation weights. Based on the process correlation weights, the baseline parameters are dynamically adjusted, and a particle filter is used to correct parameter deviations. This allows the optimized parameters to adapt in real time to changes in process parameters during the machining process, achieving dynamic optimization of process parameters, effectively suppressing the accumulation of machining errors, and improving component machining accuracy.

[0019] In terms of process anomaly monitoring and intelligent control, a multidimensional process parameter space is constructed and extreme value points and variance mutation intervals are extracted, enabling real-time monitoring of abnormal conditions during the machining process. When the extreme value density exceeds a set threshold or the variance mutation interval span exceeds a limit, it is promptly identified as a process anomaly and multi-level control instructions are generated. This enables real-time early warning and intelligent control of the machining process, reducing the risk of equipment failure and the incidence of machining quality accidents.

[0020] For the coordinated optimization of process parameters and equipment control parameters, a joint optimization model for process parameter data and equipment control parameters was established. The objective function is the sum of the weighted absolute value of process parameter deviation and control energy consumption. This approach balances energy consumption optimization and equipment load safety while ensuring processing quality. The optimal control strategy is solved using sequential quadratic programming and coupled with the optimization parameters in real time to generate a process plan that optimizes processing quality. This achieves synergy between process parameter optimization and equipment control, improving the overall performance and economic benefits of the processing system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a working principle diagram of the method for real-time optimization of process parameters of automotive parts based on machine learning according to the present invention; Figure 2 This is the design diagram of the process feature extraction module; Figure 3This is the design diagram of the principal component vector extraction module; Figure 4 This is the design diagram of the process-related weight calculation module; Figure 5 Generate a design diagram of the module for optimizing parameters. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0023] See also Figure 1-Figure 5 The present invention relates to a method for real-time optimization of process parameters of automotive parts based on machine learning, which specifically includes the following steps: A distributed sensor network collects real-time data on multi-source process parameters, including temperature gradients, pressure distribution, and cutting speed, during the machining of automotive parts. Sensors are deployed in key areas of machining equipment (such as tool contact areas, workpiece fixtures, and cooling systems), capturing dynamic data streams at nanosecond sampling rates and storing them as raw data sets in timestamp order.

[0024] A self-attention mechanism is used to extract features from process parameter data. The raw data is converted into a three-dimensional process tensor (with dimensions of time × process × parameter type), and layer normalization is used to eliminate dimensional differences. A multi-head self-attention layer calculates the global correlation weights between parameters at different time points and process stages, outputting the weighted sum of the attention weights and the feature vector to achieve cross-parameter and cross-process feature interaction. The attention results are then fed into a temporal convolutional network (TCN). Dilated convolution kernels (with exponentially increasing dilation rates) are used to extract multi-scale temporal features, capturing feature information ranging from millisecond-level fluctuations to minute-level trend changes. Gating mechanisms (such as sigmoid gates) selectively activate feature channels to suppress irrelevant noise, generating a process feature sequence that incorporates spatiotemporal correlations.

[0025] Calculate the process fluctuation coefficient based on the process feature sequence: Select the feature sequence of any process as the benchmark parameter, calculate the variance of its m adjacent process feature vectors, and traverse all processes through a sliding window. The ratio of the mean of the variance to the Manhattan distance of the benchmark parameter is defined as the intra-process fluctuation, reflecting the stability of the parameters within the same process; calculate the absolute value of the covariance between the benchmark parameter and the associated process feature as the inter-process fluctuation, representing the coupling strength of the parameters between processes. Finally, the harmonic mean of the two is used as the process fluctuation coefficient of the process, and the formula is:

[0026] Construct a covariance matrix for the process fluctuation coefficients of different processes, and obtain orthogonal basis vectors through eigenvalue decomposition (EVD). Eigenvectors whose eigenvalue variance contribution exceeds a set threshold (e.g., 85%) are selected to form a principal component subspace. The covariance matrix is ​​projected onto this subspace to achieve dimensionality reduction and obtain the principal component vectors. A nonlinear Sigmoid function is applied to the principal component vectors, compressing the values ​​to the (0, 1) range. This generates process correlation weights for each process, reflecting the strength of the parameter influence between processes.

[0027] The baseline parameters of the process feature sequence are dynamically adjusted based on process-related weights. First, the weighted adjustment parameters are obtained through matrix dot multiplication. The deviation from the baseline parameters is calculated and dynamically corrected using a particle filter (PF). Monte Carlo sampling and importance weight updates are used to suppress the effects of data noise and model uncertainty. The corrected deviations are then added to the baseline parameters to generate optimized parameters. Finally, the optimized parameters are converted into equipment control signals via Industrial Ethernet (such as PROFINET), enabling real-time adjustment of operating parameters such as the spindle speed, feed rate, and cooling flow rate of the machining equipment.

[0028] The present invention will be further described below in conjunction with Examples 1 to 5: Example 1

[0029] The process feature extraction step is implemented as follows: data preprocessing begins by sampling the raw process parameter data at equal intervals at a preset sampling frequency (e.g., 1000 Hz). This sampling frequency can be adjusted based on the response speed of the machining equipment and the frequency of parameter changes to ensure that the acquired signals accurately reflect the dynamic characteristics of the machining process. The captured data is then organized into a three-dimensional process tensor, defined as [T × N × D], where T represents the time step, corresponding to the length of the data acquisition time series; N is the number of processes, i.e., the number of different steps involved in the automotive parts manufacturing process; and D represents the parameter dimension, encompassing three parameter types: temperature gradient, pressure distribution, and cutting speed. This tensor structure enables a structured representation of heterogeneous process parameters from multiple sources across the three dimensions of time, process, and parameter type, facilitating subsequent feature extraction.

[0030] In order to eliminate the dimensional differences of different parameters, the process tensor needs to be normalized. Specifically, the data in each parameter dimension is normalized so that its mean is 0 and its standard deviation is 1. The formula for normalization is:

[0031] Here, Χ is the raw data value, μ is the mean of that parameter dimension, and σ is the standard deviation. This process converts parameters with different dimensions (such as temperature in degrees Celsius and pressure in Pascals) into dimensionless values, preventing the impact of dimensional differences on subsequent feature extraction algorithms and ensuring that all parameters have equal importance in the feature learning process.

[0032] A multi-head self-attention layer is used for global feature association. The multi-head self-attention mechanism is the core component of the Transformer architecture. Its essence is to calculate the feature association of different subspaces in parallel through multiple independent attention heads, thereby capturing richer feature interaction patterns. The specific operation process is as follows: First, the standardized process tensor is mapped into Query (query vector), Key (key vector) and Value (value vector) matrices through three linear transformation matrices. The dimensions of these three matrices are [T×N×d_k], where d k is the dimension of the key vector. Then, the attention scores between different time points and different process parameters are calculated by dot product. The calculation formula is:

[0033] in, The % is a scaling factor used to prevent excessive dot product results from causing the Softmax function to enter the gradient saturation region. The attention score reflects the strength of the association between parameters at different locations. After being normalized by the Softmax function, it is multiplied by the Value matrix to obtain the weighted attention sum, which incorporates the correlation information of all parameters globally. The multi-head mechanism uses multiple attention heads (e.g., 8 heads) in parallel, each learning a different feature correlation pattern. The outputs of each head are then concatenated and linearly transformed to produce an output containing multi-dimensional feature correlations.

[0034] After processing by the multi-head self-attention layer, the attention results are fed into a temporal convolutional network (TCN) for temporal feature extraction. TCNs employ dilated convolution kernels. The core idea behind these kernels is to insert holes between the weights of a standard convolution kernel, thereby expanding the kernel's receptive field without increasing the number of parameters. The initial dilation rate is set to 1, and after every two layers, the dilation rate increases exponentially (e.g., 1, 2, 4, 8, etc.), enabling the network to capture multi-scale temporal features, from short to long timescales. For example, a convolution kernel with a dilation rate of 1 can capture local features between adjacent time steps, while a convolution kernel with a dilation rate of 8 can capture long-range dependencies across eight time steps. In this way, TCNs can effectively extract features ranging from millisecond-level fluctuations to minute-level trend changes.

[0035] In temporal convolutional networks, gating mechanisms (such as the gated linear unit (GLU)) are also introduced to filter feature channels. A GLU consists of two parallel convolutional layers. One convolutional layer generates a gating signal using a sigmoid activation function to control whether features are passed or not; the other convolutional layer directly performs a linear transformation on the input features. The outputs of these two layers are fused through element-wise multiplication, using the formula:

[0036] Among them, W1 and W2 are convolution kernel weights, σ ​​is the Sigmoid activation function, represents element-wise multiplication. Through this gating mechanism, the network automatically suppresses irrelevant noise and selectively retains features valuable for process optimization, thereby generating a process feature sequence with spatiotemporal correlations. This sequence not only contains the numerical information of each process parameter at different time points, but also implies the global correlation between parameters and multi-scale temporal variation patterns, providing a high-quality feature representation for subsequent calculation of process fluctuation coefficients and generation of process correlation weights.

[0037] In the specific computational implementation, the parameters of the multi-head self-attention layer and the temporal convolutional network are learned and optimized using the backpropagation algorithm on training data. The training data is derived from the process parameter records of the historical machining process. By labeling the feature sequence under normal machining conditions as the target output, and using the mean squared error (MSE) as the loss function, the network weight parameters are optimized, enabling the network to accurately extract key features reflecting the machining state.

[0038] Furthermore, to ensure computational efficiency, batch processing can be employed in practical applications, combining the process tensors of multiple samples into a single batch for parallel computation. Furthermore, hardware accelerators such as GPUs can be used to optimize matrix operations, ensuring that the feature extraction process can run smoothly in machining environments with high real-time requirements. Example 2

[0039] The specific steps for calculating the process fluctuation coefficient are as follows: Define the selection rules of the benchmark parameters. For a processing flow containing n processes, select the process feature sequence of the i-th process (i∈[1,n]) As a benchmark parameter, where T is the total number of time steps, represents the d-dimensional feature vector of the i-th process at time t (d corresponds to the characteristic dimension of temperature gradient, pressure distribution, and cutting speed). The selection of benchmark parameters can be determined based on the key processes in the machining process, such as the cutting process that has the greatest impact on part accuracy, or dynamically switch benchmark processes based on real-time fluctuations to adapt to the characteristics of different machining stages.

[0040] Next, calculate the intra-process fluctuation. For the benchmark process i, traverse its m adjacent processes (m is an odd number, such as m=5, corresponding to processes i-2 to i+2, and the effective range is taken at the boundary, such as i=1 to i+2), and construct a local process set (j=(m-1) / 2). For each time t, calculate the benchmark process feature vector and adjacent process feature vectors Joint variance of (j∈local process set and j≠i) , the variance is obtained by calculating the variance of each dimension of the eigenvector and taking the mean value, the formula is:

[0041] in, represents the k-th dimension eigenvalue of benchmark process i at time t, is the sample variance function. Take the mean of the variances of all adjacent processes and time steps to get the average variance , and then compare it with the Manhattan distance of the benchmark parameter (Right now ) to obtain the intra-process fluctuation:

[0042] In the formula is a minimum value (such as 10 -8 ), to avoid situations where the denominator is zero. This calculation converts the variance into a relative fluctuation index through standardization, eliminating the influence of the absolute value of the eigenvalue on the fluctuation assessment and facilitating horizontal comparisons between different processes.

[0043] The calculation of inter-process fluctuation is based on the characteristic covariance of the benchmark process and the associated process. For any non-adjacent process k (k∉local process set), calculate the benchmark process characteristic sequence F i and process k feature sequence F k The covariance matrix of , whose elements represents the covariance between the p-th dimension feature of the benchmark process and the q-th dimension feature of process k. The Frobenius norm of the covariance matrix (i.e., the square root of the sum of the squares of each element) is taken as a comprehensive measure of the fluctuation between processes. The formula is:

[0044] in, This represents the time series of the p-th dimension of the benchmark process. By taking its absolute value (here, non-negative by taking the square root of the sum of squares), we ensure that the fluctuation between processes reflects only the strength of the association, not the positive or negative correlation.

[0045] The final process fluctuation coefficient is calculated by the harmonic mean of the intra-process fluctuation and the inter-process fluctuation. The formula is:

[0046] The harmonic mean is used to balance the impact of the two types of fluctuations, preventing one type of fluctuation from dominating the final result due to its larger value. For example, when the intra-process fluctuation is small but the inter-process fluctuation is large, the harmonic mean will tend to be smaller, indicating that the overall fluctuation in the process is not caused by a single factor, but rather by complex cross-process coupling.

[0047] In the calculation implementation, the variance and covariance are solved using a recursive algorithm to reduce the amount of calculation. For each newly collected time step data, the mean and variance in the sliding window are updated in real time to avoid repeated calculation of historical data. For example, for the variance calculation at time t, the mean of the previous time t-1 is used. and variance , the recursive formula is:

[0048] in is the eigenvalue at the current moment. This incremental calculation method reduces the time complexity from downgraded to , meeting the computational efficiency requirements of real-time optimization systems.

[0049] Furthermore, for multi-dimensional features, the overall variance and covariance of the feature vectors are used rather than single-dimensional metrics to ensure that fluctuation assessment encompasses the coupled effects of temperature, pressure, and cutting speed. For example, a temperature increase at a given moment may be accompanied by a pressure decrease. A single-dimensional variance may not reflect this negatively correlated fluctuation pattern. However, the vector-based joint variance can capture compensatory changes between features and more realistically reflect process stability.

[0050] In terms of parameter setting, the value of the number of adjacent processes, m, must be determined based on the process dependencies within the manufacturing process. For assembly-line processing (e.g., a continuous stamping-cutting-welding process), m can be set to 3-5 to account for the impact of the immediate preceding and subsequent processes. For discrete processing (e.g., multi-station parallel processing), m can be appropriately increased to include more potentially dependent processes. A dynamic switching mechanism for the benchmark process can be implemented by setting a fluctuation coefficient threshold. When the fluctuation coefficient of the current benchmark process exceeds a preset threshold (e.g., 1.5), the process with the lowest fluctuation is automatically switched to serve as the new benchmark, avoiding evaluation bias caused by inherent instability in the benchmark process. Example 3

[0051] The process of extracting the principal component vector and calculating the process-related weight is as follows: Taking the five-step manufacturing process for an automotive part as an example (Steps 1 to 5 are blank heating, stamping, cutting, surface treatment, and quality inspection, respectively), we first construct a process fluctuation coefficient matrix for each step. Assume that, using the calculation method in Example 2, we obtain the fluctuation coefficient sequence for each step over 1000 consecutive time steps (e.g., a 10-minute processing time, 100 time points per second). This results in a 5×1000 matrix C, where each row corresponds to the fluctuation coefficient time series for a single step.

[0052] Next, we calculate the covariance matrix Σ of this matrix. The covariance matrix measures the correlation between the fluctuation coefficients of different processes. For example, the temperature fluctuations of process 1 (blank heating) may be positively correlated with the pressure fluctuations of process 2 (stamping) (if the heating temperature is insufficient, the stamping pressure needs to be increased), while the cutting speed fluctuations of process 3 (cutting) may have no significant correlation with process 1. Each element Σ(i, j) of the covariance matrix represents the covariance value of the fluctuation coefficients of process i and process j. A positive value indicates that the fluctuation trends of the two processes are consistent, while a negative value indicates that the trends are opposite. The larger the absolute value, the stronger the correlation.

[0053] The covariance matrix Σ is processed through eigenvalue decomposition (EVD) to obtain a set of eigenvalues ​​and corresponding eigenvectors. An eigenvalue reflects the proportion of the variance in the fluctuation coefficient explained by the corresponding eigenvector. Larger eigenvalues ​​indicate more significant fluctuation information in that direction. Assuming that the decomposition yields five eigenvalues ​​(λ1 ≥ λ2 ≥ λ3 ≥ λ4 ≥ λ5), and the cumulative variance contribution of the first two eigenvalues, λ1 and λ2, reaches 85% (a set threshold), the eigenvectors v1 and v2 corresponding to these two eigenvalues ​​are selected to form the principal component subspace. These two eigenvectors, known as principal component vectors, are linear combinations of the original process fluctuation coefficients and are able to preserve the fluctuation information of the original data to the greatest extent possible.

[0054] Taking specific numerical values ​​as an example, assuming that the components of eigenvector v1 are [0.6, 0.7, 0.2, -0.1, 0.1] and the components of v2 are [-0.2, 0.1, 0.8, 0.5, -0.1], the larger components of process 1 and process 2 in v1 indicate that the fluctuations of these two processes are highly synergistic, possibly influenced by common factors in the same processing stage (such as early forming). The larger components of process 3 and process 4 in v2 reflect the strong correlation between the fluctuations of the cutting and surface treatment processes, which may be related to equipment switching or process parameter integration.

[0055] The covariance matrix Σ is mapped to the principal component subspace (the two-dimensional space spanned by v1 and v2) to obtain the reduced principal component vector. The principal component vector for each time step can be expressed as a linear combination of the original process fluctuation coefficient and the eigenvector. For example, the principal component value PC1(t) at time step t is 0.6 × process 1 fluctuation coefficient (t) + 0.7 × process 2 fluctuation coefficient (t) + 0.2 × process 3 fluctuation coefficient (t) - 0.1 × process 4 fluctuation coefficient (t) + 0.1 × process 5 fluctuation coefficient (t). PC2(t) is similarly calculated using v2. The reduced principal component vector is compressed from 5 dimensions to 2 dimensions, preserving 85% of the fluctuation information while simplifying the data structure and facilitating subsequent weight calculations.

[0056] Next, process correlation weights are calculated. The principal component vectors are normalized, scaling their values ​​to the [0, 1] range to avoid the influence of numerical differences on correlation calculations. The preset reference vector can be defined as a fluctuation pattern under ideal processing conditions. For example, a vector with uniform fluctuation coefficients and low correlation across all processes (e.g., [0.3, 0.3, 0.3, 0.3], assuming equal importance for all five processes), or a typical fluctuation pattern derived from historical processing data of high-quality products.

[0057] Calculate the Pearson correlation coefficient between the normalized principal component vector and the preset reference vector. This coefficient measures the degree of linear correlation between the two. For example, a correlation coefficient of 0.8 between a principal component vector and the reference vector indicates a high similarity between the current fluctuation pattern and the ideal state; a correlation coefficient of -0.5 indicates a significant difference. The correlation coefficient ranges from -1 to 1, with larger absolute values ​​indicating stronger correlation.

[0058] The correlation coefficient is input into a deep neural network (DNN) for nonlinear transformation. Assume the neural network consists of two hidden layers, with 256 neurons in the first layer and 128 neurons in the second layer. Both layers use the ReLU activation function. The output layer has five neurons (corresponding to the five processes). The Sigmoid activation function is used to compress the output values ​​to the range (0, 1) to generate initial weights. For example, an input correlation coefficient of 0.8, after calculation in the first hidden layer, may be converted into a set of intermediate values ​​containing nonlinear relationships (e.g., [1.2, -0.5, 0.9, 0.3, -0.1]). After processing in the second hidden layer and the output layer, the initial weight vector is [0.7, 0.6, 0.2, 0.3, 0.4]. Processes 1 and 2 have higher weights, indicating that their fluctuation patterns have a greater impact on the overall process.

[0059] Because the fluctuation coefficient during real-time processing may contain high-frequency noise, the initial weights need to be smoothed using a time sliding window. Assuming a sliding window size of 50 time steps (i.e., 5 seconds), the initial weights for each process are averaged within the window to obtain the smoothed process-related weights. For example, the initial weights for process 1 are 0.7, 0.68, 0.72, …, and 0.71 over 50 consecutive time steps. After smoothing, the output weight is a stable value of 0.7 ± 0.01, avoiding drastic changes in weight due to transient fluctuations.

[0060] The physical significance of process correlation weights lies in quantifying the relative impact of fluctuations in each process on overall process stability. For example, if the weight of process 1 (blank heating) is 0.7 and that of process 2 (stamping) is 0.6, these fluctuations significantly impact subsequent process parameter optimization, and their parameter adjustments warrant priority attention. On the other hand, the weight of process 3 (cutting) is only 0.2, indicating that its fluctuations are relatively independent and less relevant to the overall process, making it a secondary factor to consider during optimization.

[0061] In engineering implementation, eigenvalue decomposition and neural network calculations can be implemented using matrix operation libraries (such as NumPy) and deep learning frameworks (such as TensorFlow), leveraging GPU acceleration to improve real-time computing efficiency. Sliding window filtering can be implemented using a circular buffer, updating the mean within the window with each new data input, ensuring computational latency within milliseconds, meeting the responsiveness requirements of real-time optimization systems. Example 4

[0062] The specific implementation of optimization parameter generation and process anomaly monitoring is as follows: Take the four-step machining of an automotive gearbox housing as an example (the steps include milling, boring, tapping, and deburring). Assume that the process association weights for each step, as obtained in Example 3, are [0.8, 0.7, 0.3, 0.2], corresponding to milling, boring, tapping, and deburring, respectively. The benchmark parameter matrix F represents the characteristic sequence of each step under the standard process. For example, the benchmark parameters for the milling step include time series data for temperature gradient (30-50°C), pressure distribution (80-120 MPa), and cutting speed (100-150 m / min), with dimensions of [2000 × 4 × 3] (time step × step × parameter).

[0063] First, weight adjustment parameters are calculated. The process-related weights are dot-multiplied with the baseline parameter matrix. This means that the parameter values ​​for each time step of each process are multiplied by the corresponding weight. For example, if the temperature gradient baseline value for the milling process (weight 0.8) is 40°C, it will be adjusted to 40 × 0.8 = 32°C. Similarly, the pressure distribution baseline value for the boring process (weight 0.7) is 100 MPa, which will be adjusted to 70 MPa. This operation allows process parameters with high association weights to receive a larger adjustment margin during optimization, reflecting their dominant influence on the overall process.

[0064] Calculate the deviation between the weighted adjustment parameters and the baseline parameters—that is, the difference between the adjusted parameters and the original baseline values. For example, the deviation for milling temperature is 32-40°C (-8°C), and the deviation for boring pressure is 70-100°C (-30 MPa). Because weight adjustment is based solely on static correlations and does not account for real-time noise and model uncertainty, a particle filter is used to dynamically correct these deviations.

[0065] The particle filter implementation process is as follows: 1000 particles are initialized, each representing a possible deviation state. The initial states are uniformly distributed within ±10% of the baseline parameter (for example, the initial particle range for milling temperature deviation is -4°C to +4°C). For each new time step, the next particle state is predicted based on the system dynamic model (assuming the deviation changes slowly over time and follows a Gaussian distribution). For example, if the deviation at the previous moment was -8°C, the current deviation is predicted to be within the range of -8°C ±2°C. Then, the importance weight of each particle is calculated using the observation model. The observation value is the difference between the actual weight-adjusted parameter and the real-time parameter collected by the sensor. For example, if the real-time milling temperature is 35°C, the weight-adjusted parameter is 32°C, and the observed deviation is 3°C, the particle with the closest deviation to the particle prediction (for example, a predicted deviation of -5°C corresponds to an adjusted temperature of 35°C) is given a higher weight. A resampling step retains high-weighted particles and discards low-weighted particles. The mean of the remaining particles is ultimately used as the corrected deviation. For example, after correction, the milling temperature deviation is adjusted from -8°C to -5°C, and the boring pressure deviation is adjusted from -30MPa to -25MPa.

[0066] The corrected deviation is added to the baseline parameters to generate optimized parameters. For example, the optimized milling temperature is 40 + (-5) = 35°C, and the optimized boring pressure is 100 + (-25) = 75 MPa. The optimized parameters are transmitted to the processing equipment via Industrial Ethernet, and the milling machine's cooling flow (affecting temperature) and the boring machine's feed rate (affecting pressure) are adjusted in real time to ensure that the actual processing parameters approach the optimized values.

[0067] In the process anomaly monitoring phase, a multidimensional process parameter space is constructed based on the optimized parameters. Taking the milling process as an example, the three parameters of temperature, pressure, and cutting speed are mapped to a three-dimensional coordinate system according to the processing sequence. The parameter value at each time point corresponds to a point in space, forming a distribution point set containing 2000 points. The kernel density estimation method is used to spatially reconstruct these points. The point density of each area is calculated using a smoothing function to identify dense and sparse areas of parameter distribution. For example, if the point density near the temperature of 35°C, the pressure of 75MPa, and the cutting speed of 120m / min is significantly higher than that in other areas, it indicates that this parameter combination is a typical state for normal processing.

[0068] The variance gradient and skewness coefficient of the reconstructed space are calculated. The variance gradient reflects the change in the dispersion of the parameter distribution. A sudden increase in variance in a certain area (e.g., a sudden increase in temperature fluctuation from ±2°C to ±10°C) may indicate equipment failure or material abnormality. The skewness coefficient measures the symmetry of the distribution. A left-skewed cutting speed distribution (an increase in the proportion of low-speed values) may indicate tool wear. Noise is suppressed in the gradient distribution using a Gaussian filter algorithm. For example, a 3×3 Gaussian kernel is used to smooth the point density matrix to eliminate isolated outliers caused by occasional sensor false touches or electromagnetic interference.

[0069] When the density of extreme points exceeds a set threshold (e.g., 50 points per cubic unit) or the span of a variance mutation exceeds a limit (e.g., three times the standard deviation), a process abnormality is identified. For example, if the density of extreme points (>50°C or <20°C) in milling temperature reaches 60 per cubic unit within a certain time interval, or if the variance of the pressure distribution suddenly increases from 5 MPa to 20 MPa, the system generates multi-level control instructions. The first level is a warning, alerting the operator with a flashing indicator light; the second level is an automatic fine-tuning command, increasing the cooling flow by 10% to reduce the temperature; and the third level is an emergency shutdown command, which prevents equipment damage or scrap when the variance mutation persists for more than 10 time steps.

[0070] In engineering applications, particle filter calculations can be accelerated in parallel using a field-programmable gate array (FPGA), ensuring microsecond-level correction. The construction of a multidimensional parameter space and anomaly detection utilize the parallel computing capabilities of a graphics processing unit (GPU), rendering parameter distributions and updating detection metrics in real time. Control commands are transmitted via the PROFINET industrial bus, keeping communication latency to less than 1ms, meeting the requirements for real-time control of the machining process. Example 5

[0071] The joint optimization model of process parameters and equipment control parameters is constructed as follows: Taking the three-step machining of an automotive crankshaft as an example (forging, rough turning, and fine grinding), the objective function is defined as the weighted absolute value of the process parameter deviation and the control energy consumption. The process parameters include forging temperature, rough turning depth of cut, and fine grinding feed rate. The control parameters correspond to the motor power of the forging press, the spindle speed of the rough turning lathe, and the hydraulic pump pressure of the fine grinding machine. Assuming the ideal process parameters are forging temperature of 1100°C, rough turning depth of cut of 2mm, and fine grinding feed rate of 0.5m / min, the objective function should minimize the deviation between the actual parameters and the ideal values ​​while reducing equipment energy consumption.

[0072] Constraints are set based on the equipment's safe operating range. For example, forging press motor power must not exceed 120% of its rated value of 200 kW (i.e., ≤ 240 kW), roughing lathe spindle speed must not exceed 3000 rpm, and fine grinding machine hydraulic pump pressure must not exceed 15 MPa. These constraints are protected by both hardware limits and software thresholds to ensure that the equipment remains within its safe operating range during optimization.

[0073] The optimal control strategy is solved using sequential quadratic programming (SQP). The objective function is first transformed into a constrained optimization problem, which is decomposed into a control parameter subproblem and a process parameter subproblem. In the control parameter subproblem, the process parameters are fixed at their current values, and the equipment control parameters are optimized to minimize energy consumption. For example, assuming the current forging temperature is 1080°C (20°C lower than the ideal value), the temperature needs to be increased by adjusting the motor power of the forging press while ensuring that the power does not exceed 240kW. In the process parameter subproblem, the control parameters are fixed, and the process parameter settings are optimized to reduce deviations from the ideal values. For example, when the motor power is fixed at 220kW, the maximum achievable forging temperature is calculated, and the rough turning and fine grinding parameters are adjusted to compensate for deviations in the preceding process.

[0074] The two subproblems are solved alternately and iteratively until convergence. For the forging process, for the first iteration, the control parameters are assumed to be the initial values ​​(motor power of 200 kW). The calculated forging temperature at this point is 1050°C, the deviation is 50°C, and the energy consumption is 200 kW·h. In the control parameter subproblem, increasing the power to 230 kW (close to the upper constraint limit) results in a calculated temperature increase of 1120°C, a decrease in the deviation to -20°C, and an increase in energy consumption to 230 kW·h. The weighted sum of the deviation and energy consumption at different power levels is compared, and the optimal value (e.g., 220 kW power, 1100°C, and 220 kW·h) is selected as the new control parameter. Subsequently, in the process parameter subproblem, the roughing depth of cut is adjusted to 1.8 mm based on the new temperature value (to compensate for thermal expansion of the material due to the temperature increase), while the fine grinding feed rate is maintained at 0.5 m / min, forming a new process parameter combination.

[0075] The iterative process is repeated, with each iteration using the BFGS algorithm to update the Lagrange multiplier and adjust the weights of the objective function and constraints, until the objective function value changes within a preset accuracy (e.g., 0.1%) between two consecutive iterations. Upon convergence, the optimal control strategy is obtained: a forging press power of 220 kW, a roughing lathe spindle speed of 2800 rpm, and a finishing mill hydraulic pump pressure of 14 MPa. The corresponding process parameters are a forging temperature of 1100°C, a roughing depth of cut of 1.9 mm, and a finishing feed rate of 0.5 m / min, all of which meet all equipment constraints.

[0076] The optimal control strategy is coupled with optimized parameters in real time to generate a comprehensive process plan. For example, the optimized parameters for the forging process (temperature 1100°C) and control instructions (power 220kW) are synchronously transmitted to the machine controller via a real-time operating system (RTOS). Based on real-time feedback from the temperature sensor, the controller dynamically maintains power output through PID control, ensuring that the temperature remains stable within the target value within ±5°C. During the rough turning process, the cutting depth is automatically adjusted to 1.9mm based on the actual workpiece dimensions after forging (obtained through online measurement). The spindle speed is dynamically fine-tuned based on cutting force sensor data, maintaining a constant speed of 2800 rpm ±50 rpm.

[0077] During the joint optimization process, the weighting coefficients of the objective function can be dynamically adjusted according to the processing stage. For example, during the pilot production phase of a new product, priority is given to ensuring process parameter accuracy (increasing the weight of the deviation term), while allowing for a moderate increase in energy consumption. During mass production, the system switches to energy consumption-prioritized mode (increasing the weight of the energy consumption term), reducing equipment operating load by slightly relaxing parameter accuracy (for example, allowing a ±10°C temperature fluctuation). This dynamic weighting mechanism enables the system to adapt to different production goals and process requirements.

[0078] During the project implementation, the joint optimization model was implemented in the C++ programming language and integrated into an industrial control computer (IPC). Multithreading technology was used to parallelize the subproblems of multiple processes. A real-time coupling module communicated with the equipment controller via the OPC UA protocol, ensuring microsecond-level synchronization of control instructions and process parameters. Historical optimization data was stored in an SQL database, enabling process engineers to analyze the parameter coupling relationships between each process, optimize preset reference vectors and weighting coefficients, and continuously accumulate process knowledge.

[0079] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0080] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A real-time optimization method for automobile parts process parameters based on machine learning, characterized in that: include: Collecting multi-source process parameter data during the machining of automotive parts, wherein the process parameter data includes temperature gradient, pressure distribution and cutting speed; The self-attention mechanism is used to extract process features from process parameter data to obtain a process feature sequence, and the process fluctuation coefficient is calculated based on the variance change of the process feature sequence; The covariance matrix of the process fluctuation coefficients of different processes is subjected to eigendecomposition to extract the principal component vector, which is then transformed nonlinearly using the Sigmoid function to obtain the process correlation weight of each process. The baseline parameters of the process feature sequence are dynamically adjusted according to the process association weights to generate optimization parameters, and the operating parameters of the processing equipment are regulated based on the optimization parameters.

2. The method for real-time optimization of automobile parts process parameters based on machine learning according to claim 1, characterized in that: The process feature extraction includes: Intercept process parameter data at a preset sampling frequency, construct process tensors and perform standardization; A multi-head self-attention layer is used to perform global feature association on the process tensor and output the weighted sum of the attention weight and the feature vector; The attention results are input into the temporal convolutional network, multi-scale temporal features are extracted by dilating the convolution kernel, and the feature channels are screened through the gating mechanism to generate a process feature sequence.

3. The method for real-time optimization of automobile parts process parameters based on machine learning according to claim 1, characterized in that: The process fluctuation coefficient calculation includes: Select the process feature sequence of any process as the benchmark parameter and calculate the variance of its m adjacent process feature vectors; The ratio of the mean of the variance to the Manhattan distance of the benchmark parameter is taken as the intra-process fluctuation; Calculate the covariance between the benchmark parameter and the associated process characteristics, and take its absolute value as the inter-process fluctuation; The harmonic mean of the intra-process fluctuation and the inter-process fluctuation is taken as the process fluctuation coefficient.

4. The method for real-time optimization of automobile parts process parameters based on machine learning according to claim 1, characterized in that: The principal component vector extraction includes: Construct the covariance matrix of process fluctuation coefficients of different processes and perform eigenvalue decomposition to obtain orthogonal basis vectors; Select the eigenvectors whose eigenvalue variance contribution rate exceeds the set threshold to form the principal component subspace; The covariance matrix is ​​mapped to the principal component subspace to obtain the principal component vector after dimensionality reduction.

5. The method for real-time optimization of automobile parts process parameters based on machine learning according to claim 1, characterized in that: The process association weight calculation includes: Normalize the principal component vector and calculate the Pearson correlation coefficient between it and the preset reference vector; The correlation coefficient is input into the deep neural network and the initial weight is generated after transformation by the hidden layer and activation function; The initial weights are smoothed and filtered through a time sliding window to output the process-related weights of each process.

6. The method for real-time optimization of automobile parts process parameters based on machine learning according to claim 1, characterized in that: The optimization parameter generation includes: Perform matrix dot multiplication of the process-related weight and the benchmark parameter to obtain the weight adjustment parameter; Calculate the deviation between the weight adjustment parameter and the benchmark parameter, and dynamically correct the deviation through the particle filter; The corrected deviation is added to the baseline parameters to generate the optimized parameters.

7. The method for real-time optimization of automobile parts process parameters based on machine learning according to claim 1, characterized in that: Also includes: Construct a multidimensional process parameter space based on the optimized parameters, and extract the extreme points and variance mutation intervals of the parameter space; When the extreme point density exceeds the set threshold or the variance mutation interval span is greater than the limit value, it is determined to be a process abnormal state and multi-level control instructions are generated.

8. The method for real-time optimization of automobile parts process parameters based on machine learning according to claim 7, characterized in that: The multi-dimensional process parameter space construction includes: Mapping the optimization parameters to a multi-dimensional coordinate system according to the processing sequence to generate a parameter distribution point set; The kernel density estimation method is used to spatially reconstruct the parameter distribution point set, and the variance gradient and skewness coefficient of the reconstructed space are calculated; The noise of gradient distribution is suppressed by Gaussian filtering algorithm to eliminate random interference.

9. The method for real-time optimization of automobile parts process parameters based on machine learning according to claim 1, characterized in that: Also includes: Establish a joint optimization model of process parameter data and equipment control parameters, and use sequential quadratic programming to solve the optimal control strategy; The optimal control strategy is coupled with the optimization parameters in real time to generate a process plan with the best processing quality; The joint optimization model solution includes: The objective function is defined as the sum of the weighted absolute values ​​of process parameter deviation and control energy consumption, and the constraint condition is the equipment load safety range; Perform KKT conditional transformation on the objective function and decompose it into control parameter sub-problem and process parameter sub-problem; The two sub-problems are solved alternately and iteratively until convergence, and the optimal control strategy that satisfies the constraints is output.

10. A real-time optimization system for automotive parts process parameters based on machine learning, characterized in that: include: Multi-source data acquisition module, used to obtain temperature gradient, pressure distribution and cutting speed data during the processing in real time, and build a multi-dimensional process parameter time series matrix; The process feature extraction module uses a self-attention mechanism to perform global feature association on the multi-dimensional process parameter time series matrix, captures multi-scale process features through a temporal convolutional network, and outputs a process feature sequence; The process fluctuation coefficient calculation module generates intra-process fluctuation and inter-process fluctuation according to the variance change of the process feature sequence, and calculates their harmonic mean as the process fluctuation coefficient; The covariance matrix decomposition module constructs the covariance matrix for the process fluctuation coefficients of different processes and extracts the principal component vector through eigenvalue decomposition; The weight mapping module uses the Sigmoid function to perform nonlinear transformation on the principal component vector to generate the process-related weights of each process; The parameter optimization module dynamically adjusts the baseline parameters based on process-related weights, corrects parameter deviations through a particle filter, and generates optimized parameters; The real-time control module converts the optimization parameters into control signals for the processing equipment, which are transmitted to the equipment controller via the industrial bus to complete the operation parameter regulation.

Images