High-precision machining control method and system based on deep learning

Abstract

The invention relates to the technical field of machining, and discloses a high-precision machining control method and system based on deep learning, and the method comprises the steps: collecting the operation data of a machine tool in real time through a multi-sensor array, and carrying out the data preprocessing; extracting a minimum feature subset most relevant to energy consumption from machine tool operation data by using an information bottleneck principle; constructing a lightweight energy consumption prediction model based on a sparse Bayesian learning method; the processing track is decomposed into two levels of macroscopic path planning and microcosmic speed planning, and linkage adjustment of the macroscopic path planning and the microcosmic speed planning is achieved through a collaborative optimization algorithm; an improved Transform deep learning architecture is adopted to process historical trajectory data, key points and modes in a trajectory are identified through a self-attention mechanism, and trajectory parameters are dynamically adjusted; energy conservation and emission reduction in the mechanical part machining process are achieved by bringing the energy consumption index into the trajectory optimization target, and energy consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the field of mechanical processing technology, and more specifically, to a high-precision machining control method and system based on deep learning. Background Art

[0002] As modern manufacturing continues to demand higher precision, higher efficiency, and lower energy consumption, traditional machining processes for mechanical parts face numerous technical challenges. Existing CNC machine tool trajectory planning relies primarily on offline optimization, which struggles to adapt to dynamic changes during the machining process, making it difficult to balance precision and efficiency. Furthermore, existing trajectory optimization systems primarily focus on time and accuracy, neglecting energy optimization, resulting in low energy efficiency.

[0003] Traditional energy optimization methods are computationally complex and difficult to achieve real-time control. The lack of accurate identification of key energy consumption factors leads to significant trajectory redundancy during machining, resulting in wasted time, energy, and tool life. These issues are particularly prominent in the machining of high-precision mechanical parts in fields such as aerospace and precision instruments.

[0004] While several methods exist for machine tool trajectory optimization, such as trajectory planning based on particle swarm algorithms and parameter optimization based on genetic algorithms, these approaches often focus on a single objective, such as minimizing time or maximizing accuracy, and fail to comprehensively optimize energy efficiency, machining accuracy, and efficiency. Furthermore, most existing methods rely on complex mathematical models and computationally intensive optimization algorithms, making real-time optimization control difficult to achieve. While deep learning technology has achieved breakthroughs in pattern recognition, predictive modeling, and intelligent decision-making in recent years, its application in machining control remains in its infancy and has yet to fully realize its potential.

[0005] Therefore, there is an urgent need for an automated control method that can achieve coordinated optimization of energy efficiency, processing accuracy and processing efficiency, and provide green and efficient processing solutions for modern manufacturing. Summary of the Invention

[0006] The present invention provides a high-precision machining control method and system based on deep learning, which solves the technical problem in related technologies that it is difficult to balance energy efficiency, machining accuracy and machining efficiency.

[0007] The present invention provides a high-precision machining control method based on deep learning, comprising:

[0008] Collect machine tool operation data in real time through a multi-sensor array and perform data preprocessing;

[0009] The information bottleneck principle is used to extract the minimum feature subset most relevant to energy consumption from the machine tool operation data;

[0010] Based on the minimum feature subset, a lightweight energy consumption prediction model is constructed based on the sparse Bayesian learning method;

[0011] Based on the output of the energy consumption prediction model, the machining trajectory is decomposed into two levels: macro-path planning and micro-speed planning, and the coordinated adjustment of the two is achieved through a collaborative optimization algorithm;

[0012] Based on machine tool operation data, minimum feature subset, output of energy consumption prediction model, macro path planning and micro speed planning, an improved Transformer deep learning architecture is used to process historical trajectory data. The key points and patterns in the trajectory are identified through the self-attention mechanism, and the trajectory parameters are dynamically adjusted to achieve real-time optimization and smoothing of the tool trajectory.

[0013] Furthermore, the step of collecting machine tool operation data in real time through a multi-sensor array includes:

[0014] Install multiple types of sensors at key locations on CNC machine tools, including position sensors, current sensors, temperature sensors, vibration sensors, and acoustic sensors;

[0015] The timestamp alignment method is used to synchronize multi-sensor data. The collected raw data is band-pass filtered to remove DC drift and high-frequency noise.

[0016] The median filter method was used to remove outliers, and the sliding window averaging method was used to smooth the data.

[0017] Furthermore, the step of extracting the minimum feature subset most relevant to energy consumption from the machine tool operation data using the information bottleneck principle is approximately optimized using a variational information bottleneck method, including constructing an encoder network and a predictor network by minimizing the following variational upper bound:

[0018]

[0019] in represents the variational information bottleneck loss function, Indicates q φ (z|X data ) expectation, q φ (z|X data ) represents the output of the encoder network, X data represents the data features collected by the original sensor, z represents the extracted compressed feature representation; p θ (e|z) represents the output of the predictor network, e represents the energy consumption index; KL(·||·) represents the KL divergence, r(z) represents the prior distribution of the random variable z, which is set to the standard normal distribution, β IB is the weight coefficient; log represents the logarithmic function.

[0020] Furthermore, the step of constructing a lightweight energy consumption prediction model based on the sparse Bayesian learning method includes:

[0021] The total energy consumption of the machine tool is decomposed into three main components: cutting energy consumption, servo system energy consumption and standby energy consumption;

[0022] Based on the extracted key features, a sparse Bayesian regression model is constructed to predict energy consumption;

[0023] Automatic correlation is used to determine the prior, automatically identifying the most relevant features and assigning them weights greater than 0.1, while suppressing the weights of irrelevant features to below 0.01 to achieve model sparsification.

[0024] Furthermore, the calculation formula for the total energy consumption of the machine tool is:

[0025]

[0026] Among them E total represents the total energy consumption; represents the sum from 0 to T, T represents the total time; Δt represents the time step; P standby (t) represents standby energy consumption; P cut (t) represents the cutting energy consumption at time t; P servo (t) represents the energy consumption of the servo system.

[0027] Furthermore, in the step of decomposing the machining trajectory into two levels: macroscopic path planning and microscopic speed planning:

[0028] Macro-path optimization uses an improved A-star algorithm based on deep learning to construct the initial path, and then uses the elastic band algorithm to smooth and optimize the path;

[0029] Micro-speed optimization uses a model predictive control method enhanced by a deep neural network to calculate the optimal speed curve while satisfying the machine tool dynamic constraints and processing quality requirements.

[0030] Furthermore, the calculation formula of the self-attention mechanism is:

[0031]

[0032] Where Attention(Q, K, V) represents the output of the self-attention calculation, Q, K, and V represent the query matrix, key matrix, and value matrix respectively, T represents the transpose operator, QK T represents the matrix product of the transpose of the query matrix Q and the key matrix K, Softmax represents the function that converts the input into a probability distribution, represents the key vector dimension d kThe square root of , used for scaling;

[0033] The query matrix, key matrix, and value matrix are obtained from the input trajectory sequence through linear transformation:

[0034] Q=X input W Q ;

[0035] K=X input W K ;

[0036] V=X input W V ;

[0037] Where Q represents the query matrix, X input is the input trajectory sequence, W Q 、W K and W V The weight matrices representing the query matrix, key matrix, and value matrix respectively.

[0038] Furthermore, the step of dynamically adjusting the trajectory parameters adopts a parameter adjustment strategy based on a deep deterministic policy gradient algorithm, and the current processing state is represented as a state vector S t , the trajectory parameter adjustment is expressed as the action vector By maximizing the cumulative reward:

[0039]

[0040] where R t represents the cumulative reward starting from time t, represents the immediate reward, which is defined as the weighted sum of the energy consumption reduction and the processing efficiency improvement; represents the immediate reward at the kth time step in the future, which is the weighted sum of the energy consumption reduction and the processing efficiency improvement; γdiscount is the discount factor, ranging from 0 to 1, which is used to balance the importance of short-term rewards and long-term rewards. ∞ represents infinity, k represents the number of future time steps from the current moment, t represents the current moment, and t+k represents the moment of the kth time step in the future; represents a summation operation from 0 to infinity, which is used to calculate the sum of discounted rewards for all future time steps.

[0041] Furthermore, the data collected by the multi-sensor array include tool position, feed speed, spindle speed, current, temperature and vibration signals; the key features extracted by the information bottleneck principle include the root mean square value of spindle current, cutting speed, feed rate, cutting depth, motor current fluctuation coefficient, spindle temperature, workpiece surface temperature gradient and tool path curvature.

[0042] The present invention provides a high-precision machining control system based on deep learning, which is used to execute the above-mentioned high-precision machining control method based on deep learning, including:

[0043] Multi-sensor data acquisition unit, used to collect machine tool operation data in real time and perform pre-processing;

[0044] An information bottleneck feature extraction unit, used to extract the minimum feature subset most relevant to energy consumption from the collected data;

[0045] Sparse Bayesian energy consumption modeling unit, used to build lightweight energy consumption prediction models;

[0046] Multi-time domain trajectory decomposition unit, used to decompose the processing trajectory into two levels: macro path planning and micro speed planning;

[0047] The self-attention trajectory optimization unit is used to process historical trajectory data, identify key points and patterns in the trajectory, and dynamically adjust trajectory parameters.

[0048] The beneficial effects of the present invention are: by incorporating energy consumption indicators into trajectory optimization targets through deep learning technology, energy conservation and emission reduction in the mechanical parts processing process are achieved, and energy consumption is reduced. The combination of deep learning and information bottleneck theory enables the system to accurately identify key factors affecting energy consumption, avoiding unnecessary calculations and optimizations;

[0049] Trajectory optimization based on the self-attention mechanism of the Transformer deep learning architecture can reduce trajectory redundancy and improve processing efficiency while maintaining or improving processing accuracy. The self-attention mechanism can learn from historical trajectory data and identify key points and patterns in the trajectory;

[0050] By screening key features through the information bottleneck theory, the amount of calculation is reduced, enabling the system to achieve real-time control under conditions of limited computing resources. The multi-time domain decomposition technology decomposes the complex trajectory optimization problem into two levels, macro and micro, greatly reducing the difficulty of solving it.

[0051] The multi-time-domain trajectory decomposition and optimization method enables the system to quickly adapt to different processing conditions and workpiece material changes, improving the system's versatility and robustness. The energy consumption model constructed using the sparse Bayesian learning method has excellent generalization capabilities.

[0052] It can be directly integrated into the existing CNC system without the need for large-scale hardware modification, with low implementation cost. The system runs stably and reliably, and can meet the comprehensive requirements of industrial production for high precision, high efficiency and low energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the automated control method for high-precision machining of mechanical parts of the present invention;

[0054] Figure 2 is a flow chart of step 1 of the present invention;

[0055] Figure 3 is a flow chart of step 2 of the present invention;

[0056] Figure 4 is a flow chart of step 3 of the present invention;

[0057] Figure 5 is a flow chart of step 4 of the present invention;

[0058] Figure 6 It is a flow chart of step 5 of the present invention. DETAILED DESCRIPTION

[0059] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0060] At least one embodiment of the present invention discloses a high-precision machining control method based on deep learning, such as Figures 1 to 6 As shown, the following steps are included:

[0061] Step 1: Real-time collection of machine tool operation data through a multi-sensor array and data preprocessing;

[0062] Step 1.1, deploy the sensor array;

[0063] Multiple types of sensors are installed at key locations on CNC machine tools, including position sensors, current sensors, temperature sensors, vibration sensors, and acoustic sensors. Position sensors monitor the relative position of the tool and workpiece in real time; current sensors monitor current changes in the spindle motor and feed motor; temperature sensors monitor temperature changes in the cutting area, spindle, and guide rails; vibration sensors monitor the vibration of the machine tool structure and tool; and acoustic sensors capture the sound characteristics of the machining process.

[0064] Step 1.2, data synchronization and preprocessing;

[0065] Timestamp alignment was used to synchronize multi-sensor data, with a sampling frequency of 1kHz to capture high-frequency transients. The raw data was bandpass filtered to remove DC drift below 0.5Hz and high-frequency noise above 100Hz. Median filtering was then used to remove outliers, and sliding window averaging was used for data smoothing.

[0066] Step 1.3, data formatting and storage;

[0067] Organize the preprocessed data in a unified format to construct a multidimensional time series data structure:

[0068] D=X t , P t , V t , I t , T t , A t ;

[0069] Where D represents the dataset, X t 、P t 、V t , I t 、T t 、A t They represent the position vector, cutting parameter vector, velocity vector, current vector, temperature vector, and acceleration vector at time t, respectively. The formatted data is stored in a time series database, and a data index is established to support efficient query.

[0070] The output of this step is high-quality, standardized, multidimensional time series data, encompassing key parameters such as position, velocity, current, temperature, and acceleration during machine operation. This data provides the foundation for subsequent feature extraction and energy consumption model construction, ensuring that the entire automated control system can make optimization decisions based on authentic and accurate machine status data.

[0071] Step 2: Use the information bottleneck principle to extract the minimum feature subset most relevant to energy consumption from the machine tool operation data;

[0072] Step 2.1, information bottleneck model construction;

[0073] An information bottleneck model is constructed to extract the smallest but most informative feature subset. Based on the principles of information theory, this model achieves data compression and key information retention by maximizing the mutual information between features and energy consumption indicators while minimizing the mutual information between features and original data. The optimization objective is expressed as:

[0074] max Φ I(Φ(X data );E energy )-β IB I(Φ(Xdata );X data );

[0075] where max Φ Represents the optimization of the mapping function Φ to maximize the objective function, X data Represents the original trajectory state characteristics, E energy represents the energy consumption index, β IB is a trade-off coefficient used to control the balance between compression rate and information retention rate; I(·; ·) represents the mutual information between two random variables, and the mutual information calculation formula is:

[0076]

[0077] Where I(X;Y) represents the mutual information between random variables X and Y, represents the sum of all possible values ​​of the random variable X. represents the sum of all possible values ​​of the random variable Y, p(x, y) represents the joint probability distribution, p(x) and p(y) represent the marginal probability distributions, log represents the logarithmic function, and x and y represent random variables.

[0078] Step 2.2, implementation of variational information bottleneck algorithm;

[0079] Since direct optimization of mutual information has high computational complexity, the variational information bottleneck method is used for approximate optimization. Construct the encoder network q φ (z|X data ) and the predictor network p θ (e|z), by minimizing the following variational upper bound:

[0080]

[0081] in represents the variational information bottleneck loss function, Indicates q φ (z|X data ) expectation, q φ (z|X data ) represents the output of the encoder network, X data represents the data features collected by the original sensor, z represents the extracted compressed feature representation; p θ (e|z) represents the output of the predictor network, e represents the energy consumption index; KL(·||·) represents the KL divergence, r(z) represents the prior distribution of the random variable z, which is set to the standard normal distribution, β IB is the weight coefficient; log represents the logarithmic function.

[0082] Step 2.3, feature importance evaluation and selection;

[0083] Based on the trained variational information bottleneck model, the importance score of each feature for energy consumption prediction is calculated. A permutation-based feature importance assessment method is used to quantify the importance of a feature by randomly permuting its value and observing the degree of degradation in model prediction performance. Based on the importance scores, the top K features are selected as the key feature subset. K is determined through cross-validation and is typically 10%-20% of the original feature dimension.

[0084] The output of this step is a feature subset with significantly reduced dimensionality but rich information, encompassing the most critical features for energy consumption prediction. These features primarily include spindle current RMS value, cutting speed, feed rate, depth of cut, motor current fluctuation coefficient, spindle temperature, workpiece surface temperature gradient, and tool path curvature. This step is closely linked to Step 1. By performing information bottleneck analysis on the multidimensional time series data collected in Step 1, data dimensionality reduction and key information extraction are achieved, providing efficient feature input for subsequent energy consumption model construction.

[0085] Step 3: Based on the minimum feature subset, a lightweight energy consumption prediction model is constructed based on the sparse Bayesian learning method;

[0086] A lightweight energy consumption prediction model is constructed based on sparse Bayesian learning. The input of step 3 is the minimum feature subset extracted in step 2 using the information bottleneck principle. Although these features have significantly reduced dimensionality, they contain the most critical information relevant to energy consumption prediction. The output is the energy consumption prediction model.

[0087] Step 3.1, energy consumption decomposition modeling;

[0088] The total energy consumption of the machine tool is decomposed into three main components: cutting energy consumption, servo system energy consumption and standby energy consumption. The total energy consumption calculation formula is:

[0089]

[0090] Among them E total represents the total energy consumption; represents the sum from 0 to T, T represents the total time; Δt represents the time step; P standby (t) represents the standby energy consumption, which is generally a relatively constant value and represents the energy consumption of auxiliary equipment such as the machine tool control system and cooling system; P cut (t) represents the cutting energy consumption at time t, and the calculation formula is:

[0091] P cut (t) = F c (t)·v c (t);

[0092] Among them F c (t) represents the cutting force at time t, v c(t) represents the cutting speed at time t.

[0093] P servo (t) represents the energy consumption of the servo system, and the calculation formula is:

[0094]

[0095] in represents the sum from 1 to n, n represents the number of servo motors, i represents the index of the servo motor, I i (t) represents the current of the i-th servo motor, R i represents the resistance of the i-th servo motor, ω i (t) represents the angular velocity of the i-th servo motor, τ i，friction (t) represents the friction torque of the i-th servo motor.

[0096] Step 3.2, sparse Bayesian regression model construction;

[0097] Based on the key features extracted in step 2, a sparse Bayesian regression model is constructed to predict energy consumption. Sparse Bayesian regression, based on the automatic relevance determination (ARD) prior, can automatically identify the most relevant features and assign them weights greater than 0.1, while suppressing the weights of irrelevant features to less than 0.01, thus achieving model sparsification.

[0098] The model is represented as:

[0099]

[0100] Where E represents the predicted energy consumption value, f θ represents the model function with parameter θ, Φ(X features ) represents the feature subset X after the information bottleneck extraction in step 2 features The feature set obtained is further processed. represents the sum from 1 to K, K represents the number of features, i1 represents the index of the feature, Represents the weight parameter of the i1th feature, and the superscript model indicates that the parameter belongs to the model parameter. Indicates that the i1th extracted feature function acts on the feature subset X features The result, Represents the bias parameter of the model, and the superscript model indicates that the parameter belongs to the model parameter. The prior distribution of the weight is set to a zero-mean Gaussian distribution:

[0101]

[0102] in Indicates that at a given precision parameter Conditional weight parameters The conditional probability distribution of , N represents Gaussian distribution (normal distribution), Represents the weight parameter The mean of is 0, represents the variance, It is the precision parameter. The superscript ARD indicates that this parameter belongs to the Automatic Relevance Determination mechanism, which controls the sparsity of the weight.

[0103] The prior distribution of the accuracy parameter is set to a gamma distribution:

[0104]

[0105] in Indicates the accuracy parameter of the i1th feature The probability distribution of , Gamma represents the gamma distribution, Indicates the accuracy parameter of the i1th feature The shape parameter of is a0, and b0 represents the scale parameter of the gamma distribution of the i1-th feature.

[0106] Through the variational inference method, the model parameters are iteratively updated until convergence.

[0107] Step 3.3, model validation and calibration;

[0108] The energy consumption model's predictive performance was evaluated using cross-validation, with the root mean square error (RMSE) and mean absolute percentage error (MAPE) calculated as evaluation metrics. Multiple specialized energy consumption models were constructed for different workpiece materials and machining processes, and a model selection mechanism was designed to automatically select the most appropriate energy consumption prediction model based on the characteristics of the current machining task. Furthermore, an online calibration mechanism was designed to dynamically adjust model parameters based on real-time energy consumption data to improve prediction accuracy.

[0109] The output of this step is an efficient, lightweight energy consumption prediction model that can quickly and accurately predict energy consumption under different processing conditions based on key features. This model typically keeps prediction errors within 5% while reducing model complexity by over 60%, making it suitable for real-time operation in industrial control environments. This step is closely linked to the previous two steps: it directly uses the key features extracted in Step 2 as input, which are derived from the multidimensional time series data collected in Step 1. Furthermore, the model's training and validation processes also require the use of the original energy consumption monitoring data from Step 1 as labels. The constructed energy consumption prediction model will provide a critical basis for energy consumption assessment during subsequent trajectory optimization.

[0110] Step 4: Based on the output of the energy consumption prediction model, the machining trajectory is decomposed into two levels: macro-path planning and micro-speed planning, and the coordinated adjustment of the two is achieved through a collaborative optimization algorithm;

[0111] Step 4.1, trajectory multi-time domain decomposition;

[0112] Based on the principle of time-scale separation, the machining trajectory optimization problem is decomposed into two sub-problems with different time scales: macroscopic path planning and microscopic velocity planning. Macroscopic path planning focuses on the overall movement path of the tool on the workpiece and has a longer time scale; microscopic velocity planning focuses on the changes in velocity and acceleration along each path segment and has a shorter time scale. This decomposition can reduce the complexity of the optimization problem and improve the solution efficiency. Mathematically expressed as:

[0113]

[0114] in Represents the path parameter p path and velocity parameter v vel Minimization of E total (p path , V vel ) represents the total energy consumption, p path is the path parameter vector, which represents the movement trajectory of the tool on the workpiece; v vel is the velocity parameter vector, which represents the change in velocity and acceleration along the path.

[0115] The constraints are:

[0116] g(p path , v vel )≤0,h(p path , v vel )=0;

[0117] where g(p path , v vel )≤0 represents an inequality constraint;

[0118] h(p path , v vel )=0 represents an equality constraint.

[0119] Step 4.2, macro path optimization;

[0120] At the macro level, the tool path is optimized with the goal of minimizing total energy consumption and ensuring machining accuracy. An improved A-star algorithm (A*) based on deep learning is used to construct the initial path, which is then smoothed and optimized using the elastic band algorithm. The deep learning model learns the characteristics of historically optimized paths, providing the A-star algorithm with a more accurate heuristic function. The elastic band algorithm treats the path as a virtual elastic band with internal elastic forces and optimizes the path by minimizing the energy function:

[0121]

[0122] Among them E path (p path ) represents the path energy function, E internal (p path ) represents the internal deformation energy of the elastic band, which is used to keep the path smooth; E external (p path ) represents the external constraint energy, which is used to ensure the machining accuracy requirements; E energy-path (p path ) represents the energy-related energy item, which is used to minimize energy consumption; α path , β path and γ path They represent the weight coefficients of internal deformation energy, external constraint energy, and energy consumption-related energy terms, respectively, and are used to balance the importance of different objectives and ensure a trade-off between path smoothness, accuracy, and energy efficiency.

[0123] Step 4.3, micro speed optimization;

[0124] At the micro level, based on a given path, the tool velocity profile along the path is optimized to minimize energy consumption and maximize machining efficiency. Model Predictive Control (MPC) is used to calculate the optimal velocity profile while satisfying the machine tool's dynamic constraints and machining quality requirements. The MPC optimization problem is expressed as:

[0125]

[0126] in Indicates the speed parameter V vel Minimization, Represents from 0 to N pred -1 sum, N pred represents the length of the prediction time domain, k represents the index of the prediction step number; e k represents the predicted energy consumption of the kth step; Δv k Indicates the speed change, that is, the difference in speed between adjacent moments; Δa krepresents the acceleration change, that is, the difference in acceleration between adjacent moments. w1, w2, and w3 represent the weight coefficients of energy consumption, velocity change, and acceleration change, respectively, which are used to balance the three goals of minimizing energy consumption, velocity smoothing, and acceleration smoothing.

[0127] Constraints include:

[0128] v min ≤v k ≤v max ;

[0129] a min ≤a k ≤a max ;

[0130] j min ≤j k ≤j maX ;

[0131] where v k 、a k and j k Represents velocity, acceleration and jump respectively, v min and v max Represent the minimum speed constraint and the maximum speed constraint, respectively. min and a max Represent the minimum acceleration constraint and the maximum acceleration constraint, j min and j max represent the minimum jump constraint and the maximum jump constraint respectively.

[0132] The output of this step is an optimized two-layer trajectory solution: a macro-level optimized path and a micro-level optimized velocity profile. Macro-path optimization primarily outputs a smooth, efficient, and energy-efficient tool trajectory; micro-velocity optimization provides the optimal velocity and acceleration parameters for each point on the trajectory. This step is clearly linked to the previous step: it directly uses the energy consumption prediction model constructed in step 3 to evaluate the energy consumption of different trajectory and velocity strategies, and the model input comes from the key features extracted in step 2. At the same time, trajectory optimization also needs to consider the machine tool dynamic characteristic data collected in step 1, such as position, velocity, and acceleration limits. The optimized trajectory solution will provide the basic trajectory and parameters for the final self-attention trajectory dynamic optimization.

[0133] Step 5: Based on the machine tool operation data, the minimum feature subset, the output of the energy consumption prediction model, macro-path planning, and micro-speed planning, an improved Transformer deep learning architecture is used to process the historical trajectory data. The self-attention mechanism is used to identify key points and patterns in the trajectory, and the trajectory parameters are dynamically adjusted to achieve real-time optimization and smoothing of the tool trajectory.

[0134] An improved Transformer deep learning architecture is used to process historical trajectory data to achieve real-time optimization and smoothing of tool trajectories.

[0135] Step 5.1, self-attention mechanism construction;

[0136] A self-attention mechanism based on the Transformer architecture is designed to capture temporal dependencies and patterns in historical trajectory data. The input is a trajectory sequence containing information such as position, velocity, acceleration, and energy consumption. The self-attention mechanism calculates the correlation strength between each element in the sequence and identifies key points and patterns in the trajectory. The self-attention calculation formula is:

[0137]

[0138] Where Attention(Q, K, V) represents the output of the self-attention calculation, Q, K, and V represent the query matrix, key matrix, and value matrix respectively, T represents the transpose operator, QK T represents the matrix product of the transpose of the query matrix Q and the key matrix K, Softmax represents the function that converts the input into a probability distribution, represents the key vector dimension d k The square root of , used for scaling;

[0139] The query matrix, key matrix, and value matrix are obtained from the input trajectory sequence through linear transformation:

[0140] Q=X input W Q ;

[0141] K=X input W K ;

[0142] V=X input W V ;

[0143] Where Q represents the query matrix, X input is the input trajectory sequence, W Q 、W K and W V The weight matrices representing the query matrix, key matrix, and value matrix respectively.

[0144] Step 5.2, multi-head attention and residual connection;

[0145] To enhance the expressive power of the model, a multi-head attention mechanism is used, combined with residual connections and layer normalization. The multi-head attention calculation formula is:

[0146]

[0147] Among them, MultiHead(X input ) represents the output of the multi-head attention mechanism, head1, head2, Respectively represent the first, second, and hth heads The output of the attention head, h heads is the number of attention heads, usually set to 8; COncat represents the connection operation, which connects multiple tensors into a larger tensor in a specific dimension, W O is the output linear transformation matrix;

[0148]

[0149] in represents the output of the i2th attention head, and They represent the weight matrices of the query matrix, key matrix, and value matrix of the i2th attention head, respectively. Attention represents the self-attention calculation function. The residual connection and layer normalization calculation formulas are:

[0150] X′=LayerNorm(X input +MultiHead(X input ));

[0151] X″=LayerNorm(X′+FFN(X′));

[0152] LaverNorm represents the layer normalization operation, which is used to standardize the mean and variance of the input to improve training stability; X′ represents the output after the first layer of Transformer processing, X″ represents the output after the complete Transformer layer processing, and X input Represents the input data of the Transformer layer;

[0153] FFN represents a feedforward neural network layer, which contains two linear transformations and a ReLU activation function:

[0154]

[0155] Where FFN(x1) represents the output of the feedforward neural network, X1 represents the input feature vector, that is, the feature representation X′ after processing by the multi-head attention mechanism. represents the first weight matrix of the feedforward neural network, represents the first bias vector of the feedforward neural network, represents the second weight matrix of the feedforward neural network, represents the second bias vector of the feedforward neural network, and max(0, ) represents the ReLU activation function.

[0156] Step 5.3, dynamic adjustment of trajectory parameters;

[0157] Based on the output of the self-attention model and the prediction results of the energy consumption model, the trajectory parameters are dynamically adjusted. A parameter adjustment strategy based on reinforcement learning is designed, and the current processing state is represented as a state vector S t , the trajectory parameter adjustment is expressed as the action vector By maximizing the cumulative reward R t Learning the optimal policy:

[0158]

[0159] where R t represents the cumulative reward starting from time t, represents the immediate reward, which is defined as the weighted sum of the energy consumption reduction and the processing efficiency improvement; represents the immediate reward at the kth time step in the future, which is the weighted sum of the energy consumption reduction and the processing efficiency improvement; γ discount is a discount factor ranging from 0 to 1, which is used to balance the importance of short-term rewards and long-term rewards. ∞ represents infinity, k represents the number of future time steps from the current moment, t represents the current moment, and t+k represents the moment of the kth time step in the future. represents a summation operation from 0 to infinity, which is used to calculate the sum of discounted rewards for all future time steps;

[0160] Actor network output action:

[0161]

[0162] in represents the trajectory parameter adjustment action at time t, Denote the parameter θ actor The actor network function, s t represents the processing state vector at time t, θ actor represents the set of parameters of the actor network. The critic network estimates the state-action value function:

[0163]

[0164] in The parameter is φ critic The critic network function is used to estimate the state-action pair The value of φ critic represents the parameter set of the critic network, s t represents the processing state vector at time t, Represents the trajectory parameter adjustment action at time t. By alternately updating the parameters θ of the two networksactor and φ critc , and finally the optimal trajectory parameter adjustment strategy is obtained.

[0165] The output of this step is an integrated, real-time dynamic trajectory adjustment system. This system intelligently optimizes trajectory parameters in real time based on the current machining state and historical trajectory data, enabling adaptive control during machining. This step forms a complete closed-loop system with all previous steps: It uses the multi-sensor real-time data acquisition system in step 1 to acquire the current machine tool state; characterizes the current machining state using key features extracted in step 2; uses the energy consumption prediction model constructed in step 3 to evaluate the energy consumption performance of different trajectory parameters; and, based on the macroscopic path and microscopic velocity scheme generated in step 4, dynamically optimizes the trajectory using a self-attention mechanism and reinforcement learning methods. This step closes the entire control method, enabling full lifecycle management of the machining process, with the ability to flexibly adjust strategies based on the actual machining state to adapt to changes in materials, environment, and machining requirements. The ultimate output is a complete machining solution that is energy-efficient, precise, and efficient.

[0166] A high-precision machining control system based on deep learning, used to execute the above-mentioned high-precision machining control method based on deep learning, comprising:

[0167] Multi-sensor data acquisition unit, used to collect machine tool operation data in real time and perform pre-processing;

[0168] An information bottleneck feature extraction unit, used to extract the minimum feature subset most relevant to energy consumption from the collected data;

[0169] Sparse Bayesian energy consumption modeling unit, used to build lightweight energy consumption prediction models;

[0170] Multi-time domain trajectory decomposition unit, used to decompose the processing trajectory into two levels: macro path planning and micro speed planning;

[0171] The self-attention trajectory optimization unit is used to process historical trajectory data, identify key points and patterns in the trajectory, and dynamically adjust trajectory parameters.

[0172] Here, the present invention provides an implementation example:

[0173] To verify the effectiveness of this implementation, we conducted an application test on a CNC milling machine at a precision machinery parts manufacturer. The test object was an aircraft engine blade made of the high-temperature alloy GH4169, with a machining accuracy requirement of ±0.01mm.

[0174] The company faces the following challenges in the processing of aircraft engine blades: high-temperature alloy materials are difficult to process and tool wear is severe; the blade surface is complex and trajectory planning and optimization are difficult; the processing process consumes a lot of energy, but the existing system lacks effective energy consumption optimization methods; the production tasks are heavy, and it is necessary to improve processing efficiency while ensuring processing accuracy.

[0175] The company uses a DMG MORI DMU60 monoBlock five-axis CNC milling machine equipped with a Siemens 840D control system. We deployed the automated control method described in this implementation on this machine and, by integrating it with the existing control system, achieved real-time optimization of energy-efficient machining trajectories.

[0176] The sensor configuration used in this application example is shown in Table 1:

[0177] Table 1: Sensor layout and parameter configuration

[0178]

[0179] After applying the information bottleneck principle to extract key features related to energy consumption, we obtained the feature importance ranking, as shown in Table 2:

[0180] Table 2: Feature importance evaluation results (top 10)

[0181]

[0182] By screening using the information bottleneck principle, 14 of the most critical features were retained from the original 56 features, reducing the feature dimension by 75% while the prediction accuracy only dropped by 2.3%.

[0183] The verification results of the energy consumption prediction model built based on the sparse Bayesian learning method are shown in Table 3:

[0184] Table 3: Energy consumption prediction model performance evaluation

[0185] Artifact Type Number of test samples RMSE(kW) MAPE (%) <![CDATA[R 2 ]]> Leaf root 120 0.32 3.8 0.925 Leaf blade 160 0.29 3.5 0.938 Leaf tip 80 0.35 4.1 0.912 Whole blade 200 0.40 4.6 0.894

[0186] For typical surfaces in blade processing, we apply the self-attention mechanism to trajectory optimization. The results are shown in Table 4:

[0187] Table 4: Example data for self-attention trajectory optimization

[0188]

[0189] The key indicators of the processing before and after the application of this embodiment are compared, as shown in Table 5:

[0190] Table 5: Performance comparison before and after trajectory optimization

[0191] Evaluation indicators Traditional trajectory planning This embodiment Percent improvement Energy consumption (kWh / unit) 8.6 6.1 -29.1％ Processing time (min / piece) 45.3 36.5 -19.4％ Tool life (pieces / tools) 12 16 +33.3％ Surface roughness Ra (μm) 0.8 0.7 -12.5％ Dimensional accuracy deviation (μm) 9.5 8.2 -13.7％

[0192] We compared the performance of the multi-time-domain trajectory decomposition method under different process parameters, as shown in Table 6:

[0193] Table 6: Performance of the multi-time domain trajectory decomposition method under different process parameters

[0194]

[0195] After six months of practical application testing, this implementation method has demonstrated significant economic benefits in aircraft engine blade processing, as shown in Table 7:

[0196] Table 7: Economic Benefit Analysis

[0197] Benefit indicators Numerical Energy cost savings Annual electricity savings are approximately 126,000 kWh, saving approximately RMB 101,000 in electricity bills Improved production efficiency Annual output increased by about 850 pieces, with an increased output value of about RMB 680,000 Tool cost savings Tool consumption was reduced by approximately 25%, saving approximately RMB 43,000 in costs Total economic benefits Annual cost savings of approximately RMB 824,000 Payback period 6 months

[0198] Verified by the data from the above real-world application examples, this implementation demonstrates significant technical advantages and application value in the field of high-precision machining of mechanical parts. It not only achieves a significant reduction in energy consumption, but also improves machining efficiency and precision, extends tool life, and brings considerable economic benefits to the enterprise.

[0199] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A high-precision machining control method based on deep learning, characterized in that: include: Collect machine tool operation data in real time through a multi-sensor array and perform data preprocessing; The information bottleneck principle is used to extract the minimum feature subset most relevant to energy consumption from the machine tool operation data; Based on the minimum feature subset, a lightweight energy consumption prediction model is constructed based on the sparse Bayesian learning method; Based on the output of the energy consumption prediction model, the machining trajectory is decomposed into two levels: macro-path planning and micro-speed planning, and the coordinated adjustment of the two is achieved through a collaborative optimization algorithm; Based on machine tool operation data, minimum feature subset, output of energy consumption prediction model, macro path planning and micro speed planning, an improved Transformer deep learning architecture is used to process historical trajectory data. The key points and patterns in the trajectory are identified through the self-attention mechanism, and the trajectory parameters are dynamically adjusted to achieve real-time optimization and smoothing of the tool trajectory.

2. The high-precision machining control method based on deep learning according to claim 1, characterized in that: The step of collecting machine tool operation data in real time through a multi-sensor array includes: Install multiple types of sensors at key locations on CNC machine tools, including position sensors, current sensors, temperature sensors, vibration sensors, and acoustic sensors; The timestamp alignment method is used to synchronize multi-sensor data. The collected raw data is band-pass filtered to remove DC drift and high-frequency noise. The median filter method was used to remove outliers, and the sliding window averaging method was used to smooth the data.

3. The high-precision machining control method based on deep learning according to claim 1, characterized in that: The step of extracting the minimum feature subset most relevant to energy consumption from the machine tool operation data using the information bottleneck principle is approximately optimized using the variational information bottleneck method, including constructing an encoder network and a predictor network by minimizing the following variational upper bound: in represents the variational information bottleneck loss function, Indicates q φ (z|X data ) expectation, q φ (z|X data ) represents the output of the encoder network, X data represents the data features collected by the original sensor, z represents the extracted compressed feature representation; p θ (e|z) represents the output of the predictor network, e represents the energy consumption index; KL(·||·) represents the KL divergence, r(z) represents the prior distribution of the random variable z, which is set to the standard normal distribution, β IB is the weight coefficient; log represents the logarithmic function.

4. The high-precision machining control method based on deep learning according to claim 1, characterized in that: The steps of constructing a lightweight energy consumption prediction model based on the sparse Bayesian learning method include: The total energy consumption of the machine tool is decomposed into three main components: cutting energy consumption, servo system energy consumption and standby energy consumption; Based on the extracted key features, a sparse Bayesian regression model is constructed to predict energy consumption; Automatic correlation is used to determine the prior, automatically identifying the most relevant features and assigning weights greater than 0.1, while suppressing the weights of irrelevant features to below 0.01 to achieve model sparsification.

5. The high-precision machining control method based on deep learning according to claim 4, characterized in that: The calculation formula for the total energy consumption of the machine tool is: Among them E total represents the total energy consumption; represents the sum from 0 to T, T represents the total time; Δt represents the time step; P standby (t) represents standby energy consumption; P cut (t) represents the cutting energy consumption at time t; P servo (t) represents the energy consumption of the servo system.

6. The high-precision machining control method based on deep learning according to claim 1, characterized in that: In the step of decomposing the machining trajectory into two levels: macroscopic path planning and microscopic speed planning: Macro-path optimization uses an improved A-star algorithm based on deep learning to construct the initial path, and then uses the elastic band algorithm to smooth and optimize the path; Micro-speed optimization uses a model predictive control method enhanced by a deep neural network to calculate the optimal speed curve while satisfying the machine tool dynamic constraints and processing quality requirements.

7. The high-precision machining control method based on deep learning according to claim 1, characterized in that: The calculation formula of the self-attention mechanism is: Where Attention(Q, K, V) represents the output of the self-attention calculation, Q, K, and V represent the query matrix, key matrix, and value matrix respectively, T represents the transpose operator, QK T represents the matrix product of the transpose of the query matrix Q and the key matrix K, softmax represents the function that converts the input into a probability distribution, represents the key vector dimension d k The square root of , used for scaling; The query matrix, key matrix, and value matrix are obtained from the input trajectory sequence through linear transformation: Q=X input W Q ; K=X input W K ; V=X input W V ; Where Q represents the query matrix, X input is the input trajectory sequence, W Q 、W K and W V The weight matrices representing the query matrix, key matrix, and value matrix respectively.

8. The high-precision machining control method based on deep learning according to claim 1, characterized in that: The step of dynamically adjusting the trajectory parameters adopts a parameter adjustment strategy based on a deep deterministic policy gradient algorithm, and the current processing state is represented as a state vector s t , the trajectory parameter adjustment is expressed as the action vector By maximizing the cumulative reward: where R t represents the cumulative reward starting from time t, represents the immediate reward, which is defined as the weighted sum of the energy consumption reduction and the processing efficiency improvement; represents the immediate reward at the kth time step in the future, which is the weighted sum of the energy consumption reduction and the processing efficiency improvement; γ discount is a discount factor ranging from 0 to 1, which is used to balance the importance of short-term rewards and long-term rewards. ∞ represents infinity, k represents the number of future time steps from the current moment, t represents the current moment, and t+k represents the moment of the kth time step in the future. represents a summation operation from 0 to infinity, which is used to calculate the sum of discounted rewards for all future time steps.

9. The high-precision machining control method based on deep learning according to claim 1, characterized in that: The data collected by the multi-sensor array include tool position, feed speed, spindle speed, current, temperature and vibration signals; the key features extracted by the information bottleneck principle include the root mean square value of spindle current, cutting speed, feed rate, cutting depth, motor current fluctuation coefficient, spindle temperature, workpiece surface temperature gradient and tool path curvature.

10. A high-precision machining control system based on deep learning, characterized in that: A high-precision machining control method based on deep learning for executing any one of claims 1 to 9, comprising: Multi-sensor data acquisition unit, used to collect machine tool operation data in real time and perform pre-processing; An information bottleneck feature extraction unit, used to extract the minimum feature subset most relevant to energy consumption from the collected data; Sparse Bayesian energy consumption modeling unit, used to build lightweight energy consumption prediction models; Multi-time domain trajectory decomposition unit, used to decompose the processing trajectory into two levels: macro path planning and micro speed planning; The self-attention trajectory optimization unit is used to process historical trajectory data, identify key points and patterns in the trajectory, and dynamically adjust trajectory parameters.

Images