Laser etching precision control method and system

Abstract

The invention relates to the technical field of machining precision control, in particular to a laser carving precision control method and system.The laser carving precision control system comprises a feature collecting unit, a model building and analyzing unit, a dynamic threshold value adjusting unit and an online incremental learning unit, and the feature collecting unit collects vibration, current and temperature data through a multi-source sensor array; the model construction analysis unit realizes dynamic prediction of processing parameters by combining a bidirectional long-short-term memory network with an attention mechanism, and the dynamic threshold adjustment unit dynamically updates parameters of a numerical control system based on a material hardness real-time detection and thermal coupling model. The online incremental learning unit automatically generates training samples through error data, continuously optimizes model parameters and constructs a'data acquisition-intelligent modeling-dynamic compensation-model evolution 'closed loop, so that accurate prediction and adaptive adjustment of machining parameters are realized, and the adaptability of the manufacturing process to multi-variety and small-batch working conditions is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of machining precision control, and in particular to a laser engraving precision control method and system. Background Art

[0002] Machining precision control is an important technology. In modern manufacturing, with the popularization of multi-variety small-batch production models, the complexity and uncertainty of the machining process have increased significantly. Deeply mining the implicit rules in manufacturing data through artificial intelligence technology and realizing dynamic optimization of machining parameters plays a key role in improving machining precision, extending tool life and improving production efficiency.

[0003] However, existing machining precision control technologies suffer from insufficient feature mining depth and delayed dynamic response when processing nonlinear and time-varying machining data. Traditional feature extraction methods based on a single sensor cannot effectively integrate the coupled relationships between multi-dimensional data such as vibration spectrum, spindle current, and tool temperature, resulting in incomplete representation of the machining state. Furthermore, traditional time series prediction models lack the ability to adaptively adjust the weights of sudden changes in the feature sequence, making it difficult to capture the impact of transient characteristics such as tool wear and material hardness mutations during machining. This results in significant deviations between the predicted spindle speed compensation coefficients and the actual requirements. In addition, the threshold adjustment mechanism of existing systems is mostly based on fixed process parameters and lacks real-time correlation with dynamic variables such as material hardness and cutting temperature. As a result, when the material hardness fluctuates, the compensation parameter update lags, causing dimensional deviations or tool wear. This chain reaction of incomplete data representation, model prediction deviation, and delayed threshold adjustment ultimately leads to limited manufacturing process optimization and makes it difficult to meet the high-precision requirements of modern machining. To address this technical problem, we provide a laser engraving precision control method and system. Summary of the Invention

[0004] The purpose of the present invention is to provide a laser engraving precision control method and system to solve the problems raised in the above background technology, which are specifically as follows.

[0005] 1. Because traditional single-sensor feature extraction cannot integrate multi-dimensional data, resulting in incomplete representation of machining conditions, this case deploys a multi-source sensor array through a feature acquisition unit to extract energy entropy features of data such as vibration, current, and temperature. This enables a complete representation of machining conditions, providing comprehensive data support for optimization.

[0006] 2. Because traditional time series models have difficulty capturing transient features, resulting in large deviations in the compensation coefficient, this case uses a bidirectional long-short-term memory network combined with an attention mechanism in the model construction and analysis unit to adjust the weights of mutation feature nodes. This can accurately capture transient features, improve the accuracy of compensation coefficient prediction, and reduce deviations from actual demand.

[0007] To achieve the above objectives, a laser engraving precision control method and system are provided, comprising the following units: The feature acquisition unit deploys a multi-source sensor array to collect the vibration spectrum, spindle current fluctuation and tool temperature gradient data of the processing equipment in real time, and extracts the energy entropy characteristics of the 0.5-3kHz frequency band through wavelet packet decomposition; The model construction and analysis unit constructs a time series feature mining model based on a bidirectional long short-term memory network. It inputs a 15-minute continuous energy entropy feature sequence into the trained network, outputs the spindle speed compensation coefficient and feed rate compensation coefficient for the next process cycle, and uses an attention mechanism to adjust the weight distribution of mutation feature nodes. The dynamic threshold adjustment unit establishes a dynamic threshold adjustment mechanism, calculates the threshold offset based on the real-time detection results of the hardness value of the processed material, and triggers the update of the CNC system parameters when the absolute value of the deviation between the predicted compensation coefficient and the current set value exceeds the threshold offset; When the actual processing size error exceeds the process standard range, the online incremental learning unit automatically intercepts the sensor raw data 8 seconds before the error occurs, generates incremental training samples through the sliding time window algorithm, and uses the stochastic gradient descent method with momentum term to update the fully connected layer parameters of the bidirectional long short-term memory network.

[0008] As a further improvement of the present technical solution, in the model building and analysis unit, the feature preprocessing method of the time series feature mining model includes the following steps: The energy entropy feature sequence is decomposed into two channels using wavelet packets. The first channel extracts the energy distribution gradient vector in the 0.5-2kHz frequency band, and the second channel extracts the mutation pulse density in the 2-3kHz frequency band. The local linear embedding algorithm is used to perform nonlinear dimensionality reduction on the two-channel features to generate a 16-dimensional coupled feature vector. A residual connection module is added before the bidirectional LSTM input layer, and the current feature vector is weightedly fused with the previous five historical vectors and then input into the network.

[0009] As a further improvement of this technical solution, the attention mechanism optimization method of the model construction and analysis unit includes: A hybrid attention module is constructed to extract the spatial correlation weights of temporal features through a two-dimensional convolution kernel, generate a spatial attention map of feature nodes, apply temporal attention weights to the hidden layer output of the bidirectional long short-term memory network, perform Hadamard product operation with the spatial attention map after normalization using the hyperbolic tangent function, and set a dynamic activation factor to suppress the weights of pseudo-noise nodes.

[0010] As a further improvement of this technical solution, the network training method of the model construction and analysis unit further includes: In the pre-training phase of the transfer learning framework, a benchmark model is constructed using three types of processing data: aluminum alloy, titanium alloy, and composite materials. In the online deployment phase, a domain adaptation mechanism is introduced to align the feature distribution differences under different working conditions through the gradient reversal layer, and a forget gate threshold is set. When the proportion of new working condition data exceeds the forget gate threshold, the network topology reconstruction function is enabled.

[0011] As a further improvement of the present technical solution, the real-time detection method of the hardness of the processed material in the dynamic threshold adjustment unit includes: A piezoelectric film sensor is installed on the rake face of the tool to collect the fluctuation signal of the three-dimensional component of the cutting force in real time. The Hilbert-Huang transform is performed on the fluctuation signal of the three-dimensional component to extract the instantaneous frequency and the slope of the cutting force envelope. A hardness inversion model is constructed based on the material plastic strain equation.

[0012] As a further improvement of the present technical solution, the threshold offset calculation optimization method of the dynamic threshold adjustment unit includes: A material hardness-thermomechanical coupling correction model is established, and a cutting temperature compensation term is introduced. When intermittent cutting is detected, the impact factor correction is automatically enabled in combination with the cutting temperature compensation term, and an exponential smoothing filter is applied to the final threshold.

[0013] As a further improvement of the present technical solution, the dynamic threshold adjustment unit further includes an adaptive adjustment strategy: A historical compensation decision database is constructed to calculate the optimal threshold offset values ​​under different combinations of processing material hardness and cutting temperature. A deep learning algorithm is used to dynamically optimize the parameters of the threshold offset calculation formula. When the processing accuracy fails to meet the standard after three consecutive compensations, the parameter regression test mode is triggered.

[0014] As a further improvement of this technical solution, the model building and analysis unit further includes a timing prediction compensation mechanism: A Kalman filter is cascaded after the output layer of the bidirectional long short-term memory network to perform state estimation and correction on the compensation coefficient, and a compensation transfer function model is established. When the predicted compensation exceeds the physical limit, the constraint satisfaction algorithm is activated to replan the parameter trajectory.

[0015] As a further improvement of this technical solution, the dynamic threshold adjustment unit further includes a multi-objective collaborative optimization module: A Pareto frontier analysis model is constructed to weigh the balance point between machining efficiency and tool life, a dynamic optimization function is defined, and then a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem in real time.

[0016] A second object of the present invention is to provide a method for implementing the above-mentioned laser engraving precision control system, comprising the following steps: S1. Real-time data collection of processing equipment status is performed through vibration, current, and temperature sensors. Wavelet packet decomposition is used to extract energy entropy characteristics in the 0.5-3kHz frequency band to form a multi-dimensional time series feature matrix. S2: Input the 15-minute energy entropy sequence into the bidirectional LSTM model, combine residual connection and local linear embedding dimensionality reduction to generate 16-dimensional coupling features, optimize the weight distribution through the hybrid attention mechanism, and output the dynamic compensation coefficients of the spindle speed and feed rate; S3, based on the cutting force signal, invert the real-time hardness of the material and calculate the offset of the thermal-mechanical coupling correction threshold. When the compensation coefficient deviation exceeds the threshold, the CNC system is updated and an exponential smoothing filter is used to stabilize the threshold. S4: If the machining error exceeds the limit, the data before the error is intercepted for 8 seconds to generate incremental samples, the parameters of the LSTM fully connected layer are updated, the feature distributions of the new and old working conditions are aligned through the domain adaptation mechanism, and the network topology is reconstructed when the forget gate threshold is exceeded; S5. Utilize the filter to correct the compensation coefficient state, and adopt the Pareto frontier and genetic algorithm to dynamically optimize the balance point between machining efficiency and tool life.

[0017] Compared with the prior art, the present invention has the following beneficial effects: In a laser engraving precision control method and system, the feature acquisition unit deploys a multi-source sensor array to collect multi-dimensional data such as vibration, current, and temperature in real time, extracts energy entropy features through wavelet packet decomposition, and fully characterizes the processing state, solving the problem of incomplete representation of traditional single-source data. The model construction and analysis unit is based on a bidirectional long-short-term memory network and an attention mechanism to capture transient features and optimize weight distribution. It combines dual-channel decomposition and dimensionality reduction technology to enhance the depth of feature analysis. The dynamic threshold adjustment unit calculates the threshold offset based on dynamic variables such as material hardness and cutting temperature, introduces a thermomechanical coupling model and a multi-objective optimization algorithm to achieve real-time updating of compensation parameters, solves the lag problem of traditional fixed threshold adjustment, and improves processing accuracy and tool life. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a system workflow diagram of the present invention; Figure 2 Flow chart of the method of the present invention.

[0019] The meaning of each number in the figure is: 1. Feature acquisition unit; 2. Model building and analysis unit; 3. Dynamic threshold adjustment unit; 4. Online incremental learning unit. DETAILED DESCRIPTION

[0020] 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.

[0021] The present invention provides a laser engraving precision control system, please refer to Figure 1-Figure 2 As shown, it includes the following units: Feature acquisition unit 1 deploys a multi-source sensor array to collect the vibration spectrum, spindle current fluctuation and tool temperature gradient data of the processing equipment in real time, and extracts the energy entropy characteristics of the 0.5-3kHz frequency band through wavelet packet decomposition; The model construction and analysis unit 2 constructs a time series feature mining model based on a bidirectional long short-term memory network, inputs a continuous 15-minute energy entropy feature sequence into the trained network, outputs the spindle speed compensation coefficient and feed rate compensation coefficient in the next process cycle, and uses the attention mechanism to adjust the weight distribution of the mutation feature nodes.

[0022] In the model building and analysis unit 2, the feature preprocessing method of the time series feature mining model includes the following steps: The energy entropy feature sequence is decomposed into two channels by wavelet packets. Signals such as the vibration spectrum and the spindle current contain multi-band features. Different frequency bands correspond to different physical meanings. Features need to be extracted in different frequency bands to enhance the characterization capability. The first channel uses three-layer wavelet packet decomposition to cover the energy distribution gradient vector of the 0.5-2kHz frequency band of mechanical vibration. The energy distribution gradient vector is the absolute value sequence of the energy difference between adjacent frequency bands, reflecting the energy change trend of the low-frequency band. The second channel uses four-layer wavelet packet decomposition, focusing on the high-frequency mutation signal calculation to extract the mutation pulse density in the 2-3kHz frequency band. The threshold is set to 2.5 times the standard deviation of the average energy, and a feature vector is constructed. The output dimension of the first channel is: 8 frequency bands are obtained after each layer of decomposition, and an 8-dimensional vector is formed after extracting the energy gradient. The output dimension of the second channel is: 16 frequency bands are obtained after each layer of decomposition, and a 16-dimensional vector is formed after extracting the pulse density, realizing the decoupling of high and low frequency band features.

[0023] After wavelet packet decomposition, the feature dimension is 24 in total, and there is a nonlinear coupling relationship. Traditional linear dimensionality reduction is difficult to preserve the manifold structure. Nonlinear methods are needed to reduce the computational complexity and preserve the correlation between features. First, set the parameters, that is, the number of domain points k=10, which is dynamically adjusted according to the sample density, ranging from 8-15, to ensure that the local neighborhood of each point has a linear structure. The embedding dimension d=16 is an empirical value used to balance information retention and computational efficiency. The local linear embedding algorithm is used to perform nonlinear dimensionality reduction on the two-channel features, calculate the reconstruction weight matrix of each high-dimensional feature point and the neighborhood point, minimize the local reconstruction error, and solve the low-dimensional embedding coordinates through global optimization to preserve the local linear structure of the feature manifold and avoid feature confusion caused by linear dimensionality reduction. Add a residual connection module before the bidirectional LSTM input layer, and input the current feature vector and the previous 5 history vectors into the network after weighted fusion, as follows: Design residual connection module: Input: Feature vector at the current moment (16 dimensions) and the previous 5 historical moment vectors ; Weighted fusion formula: ; Among them, the weight Obtained through adaptive learning, the initial value is equal weight 0.2, and dynamically adjusted during training, is the input after weighted fusion, For the The fused feature vector is passed through a batch normalization layer and then input into a bidirectional long short-term memory network to alleviate the gradient vanishing problem.

[0024] The attention mechanism optimization methods of model building and analysis unit 2 include: A hybrid attention module is constructed to extract the spatial correlation weights of temporal features through a two-dimensional convolution kernel. There is a coupling relationship between temperature and pressure between different dimensions in the temporal features, and this relationship needs to be explicitly modeled to enhance feature expression. The input of the two-dimensional convolution structure of the two-dimensional convolution kernel is the hidden layer output of the bidirectional long short-term memory network. The convolution kernel is configured to use a 3×3 convolution kernel, a step size of 1, a padding of 1, and 16 channels to generate a feature map. The feature map is then globally averaged pooled, and dimension-level spatial weights are generated through a fully connected layer. Finally, temporal attention weights are applied to the hidden layer output of the bidirectional long short-term memory network to generate a spatial attention map of the feature nodes.

[0025] After normalization using the hyperbolic tangent function, the Hadamard product operation is performed with the spatial attention map, as follows: The importance of different time steps in time series data is different, so temporal attention needs to be dynamically allocated. A 1×1 convolution is applied to the output of the long short-term memory network to generate temporal features. The temporal features are then substituted into the formula to calculate the attention score. The temporal attention score is then used to weight the temporal attention. The weighted temporal attention and spatial attention maps are fused using the Hadamard product, and the Hadamard product is normalized by hyperbolic tangent. A dynamic activation factor is then set to suppress the weights of pseudo-noise nodes to avoid overfitting.

[0026] The network training method of the model building and analysis unit 2 also includes: Machining data of aluminum alloy, titanium alloy, and composite materials, including vibration spectrum, spindle current, and tool temperature, are collected, and tool wear and surface roughness are annotated as regression targets. A transfer learning framework is used in the pre-training phase to construct a baseline model using the three types of machining data of aluminum alloy, titanium alloy, and composite materials. The baseline model is trained and a domain adaptation mechanism is introduced in the online deployment phase. That is, a gradient reversal layer is added after the hidden layer of the bidirectional long short-term memory network. The gradient reversal layer aligns the feature distribution differences under different working conditions, allowing the model to learn domain-invariant features. A three-layer fully connected network is used as the domain classifier, and the output is the working condition category. The domain classifier and hidden layer features are then substituted into the formula to calculate the adversarial loss. When the proportion of new working condition data exceeds a certain threshold, the original baseline model structure may not be able to effectively model, and the network topology needs to be dynamically adjusted. The proportion of new working condition samples in the current batch of data is counted in real time. The new working condition is defined as a working condition with a difference of >70% from the pre-training material. A forget gate threshold is set. When the proportion of new working condition data exceeds the forget gate threshold, the network topology reconstruction function is enabled, as follows: A new hidden layer is added after the bidirectional long short-term memory network layer to learn the specific features of the new working condition. The pre-trained layer parameters are frozen, and only the newly added layers and fully connected layers are trained. The learning rate is set, and finally L1 regularization is used to prune redundant connections. The pruning rate is controlled at 10%-15% to improve network parameter utilization.

[0027] Model building and analysis unit 2 also includes a timing prediction compensation mechanism: A Kalman filter is cascaded after the output layer of the bidirectional long short-term memory network to eliminate the output noise interference and drift error of the bidirectional long short-term memory network, perform state estimation correction on the compensation coefficient, and establish a compensation transfer function model. When the predicted compensation exceeds the physical limit, the constraint satisfaction algorithm is activated to replan the parameter trajectory.

[0028] The dynamic threshold adjustment unit 3 establishes a dynamic threshold adjustment mechanism, calculates the threshold offset according to the real-time detection result of the hardness value of the processed material, and triggers the update of the CNC system parameters when the absolute value of the deviation between the predicted compensation coefficient and the current set value exceeds the threshold offset.

[0029] The real-time detection method of the hardness of the processed material in the dynamic threshold adjustment unit 3 is realized by multi-physics field coupling sensing and nonlinear signal analysis technology, specifically including: A piezoelectric film sensor is installed on the rake face of the tool. Three sets of orthogonally mounted sensor units are used to synchronously collect the three-dimensional components of the cutting force, namely the tangential force, radial force, and axial force. These are used as the fluctuation signals of the three-dimensional cutting force components to provide high-fidelity raw data for hardness testing. The Hilbert-Huang transform is performed on the fluctuation signals of the three-dimensional components. An adaptive threshold wavelet packet noise reduction algorithm is used to filter out low-frequency interference below 800 Hz caused by spindle vibration. Each component signal is decomposed into 6-8 IMF components through ensemble empirical mode decomposition. Valid modes with a correlation coefficient greater than 0.85 are selected, and the Hilbert transform is performed on the selected IMF components to extract the instantaneous frequency and the slope of the cutting force envelope. The slope is calculated using the weighted least squares method to eliminate the interference of the tool geometric parameters on the test results. A hardness inversion model is constructed based on the material plastic strain equation to realize online non-destructive testing of material hardness.

[0030] The threshold offset calculation optimization method of the dynamic threshold adjustment unit 3 includes establishing a material hardness-thermomechanical coupling correction model and introducing a cutting temperature compensation term. When intermittent cutting is detected, the impact factor correction is automatically enabled in combination with the cutting temperature compensation term, and an exponential smoothing filter is applied to the final threshold, as follows: Material hardness directly affects cutting force and heat generation. Thermal factors need to be coupled to accurately calculate the threshold offset. The coupling model architecture is: ;in, is the material hardness, which is obtained by inverting the piezoelectric film sensor signal. is the cutting temperature, measured by infrared thermal imager, = is the coupling coefficient. Intermittent cutting will generate impact loads. It is necessary to additionally correct the threshold to avoid compensation lag. The time domain waveform of the vibration signal is analyzed. When the amplitude mutation is greater than 3 times the standard deviation and the duration is less than 200ms, it is determined to be intermittent cutting. The impact factor is calculated. The calculation formula of the impact factor is: ;in, is the amplitude change, is the impact strength coefficient, The impact factor is used to quantify the correction coefficient of the impact load on the threshold offset during intermittent cutting, reflecting the degree of mutation from continuous to intermittent cutting state. is the maximum amplitude value, then the threshold correction formula is ; is the corrected threshold offset. The original threshold calculation result may contain high-frequency noise and needs to be smoothed to ensure the stability of the CNC system parameter update. The exponential smoothing formula is: ;in is the smoothing coefficient, It is the smoothing value of the previous moment. When the cutting temperature gradient is greater than 5℃ / s, it will automatically increase. To 0.5, improve the response speed, maintain under stable working conditions To suppress noise, a single model may have deviations, and it is necessary to combine multi-sensor data cross-validation to improve reliability. The hardness inversion value, cutting temperature measured value, and vibration impact signal are integrated to build a three-dimensional verification matrix. When the difference between any two source data is greater than 15%, the redundant sensor calibration is triggered. After the end of daily production, the model parameters are calibrated using the actual size error of the workpiece to update The coupling coefficient is different. Different processing stages have different sensitivity requirements for threshold adjustment and need to be dynamically adapted. For rough processing, the threshold offset fluctuation is allowed to be ±10%, giving priority to ensuring efficiency. For fine processing, the threshold offset fluctuation is controlled within ±3%, and high sensitivity correction is enabled.

[0031] The dynamic threshold adjustment unit 3 also includes an adaptive adjustment strategy: A historical compensation decision database is constructed, and historical processing data is accumulated to form a knowledge base. This provides a reference threshold for the current working conditions, and the optimal threshold offset values ​​under different combinations of processing material hardness and cutting temperature are statistically analyzed. A deep learning algorithm is used to dynamically optimize the parameters of the threshold offset calculation formula. Traditional formula parameters are difficult to adapt to changes in complex working conditions and require real-time optimization through deep learning. When the processing accuracy does not meet the standard after three consecutive compensations, the parameter regression test mode is triggered. When continuous compensation fails, the parameter problem needs to be systematically checked to avoid the continued impact of incorrect parameters on the processing quality. The current parameters are frozen and rolled back to the historical optimal parameter combination. A gradient descent search is performed to find a better solution in the parameter space. The test parameters are run in a virtual simulation environment for 10 cycles, and the accuracy is verified to be improved by ≥5% before formal application.

[0032] The dynamic threshold adjustment unit 3 also includes a multi-objective collaborative optimization module, which constructs a Pareto frontier analysis model, weighs the balance point between machining efficiency and tool life, defines a dynamic optimization function, and then uses a non-dominated sorting genetic algorithm to solve the multi-objective optimization problem in real time. The details are as follows: There is a conflict between machining efficiency and tool life. It is necessary to find a non-dominated solution set through Pareto optimal theory and define the objective function. The machining efficiency is: , material removal rate per unit time, calculated as ;in, is the feed speed, is the cutting depth, is the cutting width and the tool life is , the calculation formula is ;in, is the cutting speed, is the material correlation coefficient, and constraints are added, where the machine power constraint is ,in, , Main cutting force, surface quality constraint, surface roughness , through the empirical formula Estimate, is the tool tip arc radius, and the grid search method is used in the parameter space Generate an initial solution set internally with a resolution of 0.1, apply non-dominated sorting to screen the Pareto optimal solution, and construct the initial frontier. Different production stages have different preferences for efficiency and lifespan, so the optimization goal needs to be dynamically adjusted. The preference coefficient is introduced and the preference factor is defined. , the optimization function is ;in, Set by the production planning department based on order priority and dynamically adjusted based on tool wear status Traditional optimization algorithms are difficult to handle multi-objective, nonlinear, and high-constraint problems. Intelligent algorithms need to be used to encode chromosomes and initialize the population. 100 individuals are randomly generated to ensure uniform coverage of the parameter space. Evolutionary operations are then performed to divide the population into different levels, with individuals with high levels being retained first to maintain the diversity of the solution set and prevent the algorithm from falling into local optimality. The termination condition is that the optimal solution changes by less than 1% for five consecutive generations or the maximum number of iterations reaches 100.

[0033] The collaborative method between the model building and analysis unit 2 and the dynamic threshold adjustment unit 3 includes establishing a compensation decision credibility evaluation index. When the credibility evaluation index is less than a preset threshold, the modal switching mechanism is triggered, and a compensation strategy based on the physical model is preferentially adopted. During the bidirectional long short-term memory network update phase, the material hardness feature is injected as an auxiliary training label, as follows: It is necessary to quantify the predictive reliability of the data-driven model to avoid machining anomalies due to model failure, calculate the historical variance of the long-term and short-term memory network output compensation coefficient, reflect the model uncertainty, compare the Mahalanobis distance between the current input features and the training set features, measure the similarity of the working conditions, and check whether the compensation coefficient meets the machine tool parameter constraints. When the credibility of the data-driven model is insufficient, it is necessary to switch to the physical model to ensure machining stability. The preset threshold is 0.6. When the credibility evaluation index is less than 0.6, the switch is triggered and the compensation strategy based on the empirical formula of cutting force is enabled, that is, ;in, is the proportionality coefficient, is the measured / nominal cutting force, and then the first-order inertia link transition is adopted. ;in It decays linearly from 1 to 0 to avoid parameter mutations that impact the machine tool. Material hardness is a key variable that affects the processing state. It is necessary to strengthen the model's learning ability for this feature, normalize the material hardness value, generate a hardness change rate feature, and reflect the recent hardness fluctuation trend. A hardness alignment term is added to the loss function of the long and short-term memory network, and additional weights are assigned to hardness-related features in the attention mechanism. The model and threshold adjustment must be fed back through actual processing results to form an optimization closed loop. After each workpiece is processed, the actual dimensional error, tool wear, and processing time data are recorded. If the actual error exceeds the process standard and the switching mechanism is not triggered, the credibility threshold is updated, and the processing result data is used to correct the parameters of the cutting force empirical formula. The least squares method is used for iterative optimization to solve the limitations of a single model in complex industrial scenarios.

[0034] When the actual processing size error exceeds the process standard range, the online incremental learning unit 4 automatically intercepts the sensor raw data 8 seconds before the error occurs, generates incremental training samples through the sliding time window algorithm, and uses the stochastic gradient descent method with momentum term to update the fully connected layer parameters of the bidirectional long short-term memory network.

[0035] In the present invention, the feature acquisition unit 1 collects vibration, current, and temperature data through a multi-source sensor array to extract energy entropy features. The model construction and analysis unit 2 uses a bidirectional long short-term memory network combined with an attention mechanism to realize dynamic prediction of processing parameters. The dynamic threshold adjustment unit 3 dynamically updates the CNC system parameters based on real-time detection of material hardness and a thermomechanical coupling model. The online incremental learning unit 4 automatically generates training samples through error data, continuously optimizes model parameters, and constructs a "data acquisition-intelligent modeling-dynamic compensation-model evolution" closed loop to achieve accurate prediction and adaptive adjustment of processing parameters, significantly improving the adaptability of the manufacturing process to multi-variety and small-batch working conditions.

[0036] A second object of the present invention is to provide a method for implementing the above-mentioned laser engraving precision control system, comprising the following steps: S1. Real-time data collection of processing equipment status is performed through vibration, current, and temperature sensors. Wavelet packet decomposition is used to extract energy entropy characteristics in the 0.5-3kHz frequency band to form a multi-dimensional time series feature matrix. S2: Input the 15-minute energy entropy sequence into the bidirectional LSTM model, combine residual connection and local linear embedding dimensionality reduction to generate 16-dimensional coupling features, optimize the weight distribution through the hybrid attention mechanism, and output the dynamic compensation coefficients of the spindle speed and feed rate; S3, based on the cutting force signal, invert the real-time hardness of the material and calculate the offset of the thermal-mechanical coupling correction threshold. When the compensation coefficient deviation exceeds the threshold, the CNC system is updated and an exponential smoothing filter is used to stabilize the threshold. S4: If the machining error exceeds the limit, the data before the error is intercepted for 8 seconds to generate incremental samples, the parameters of the LSTM fully connected layer are updated, the feature distributions of the new and old working conditions are aligned through the domain adaptation mechanism, and the network topology is reconstructed when the forget gate threshold is exceeded; S5. Utilize the filter to correct the compensation coefficient state, and adopt the Pareto frontier and genetic algorithm to dynamically optimize the balance point between machining efficiency and tool life.

[0037] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A laser engraving precision control system, characterized in that: The following units are included: The feature acquisition unit (1) deploys a multi-source sensor array to collect the vibration spectrum, spindle current fluctuation and tool temperature gradient data of the processing equipment in real time, and extracts the energy entropy characteristics of the 0.5-3kHz frequency band through wavelet packet decomposition; The model construction and analysis unit (2) constructs a time series feature mining model based on a bidirectional long short-term memory network, inputs a 15-minute continuous energy entropy feature sequence into the trained network, outputs the spindle speed compensation coefficient and feed rate compensation coefficient in the next process cycle, and uses an attention mechanism to adjust the weight distribution of mutation feature nodes; The dynamic threshold adjustment unit (3) establishes a dynamic threshold adjustment mechanism, calculates the threshold offset according to the real-time detection result of the hardness value of the processed material, and triggers the update of the CNC system parameters when the absolute value of the deviation between the predicted compensation coefficient and the current set value exceeds the threshold offset; When the actual processing size error exceeds the process standard range, the online incremental learning unit (4) automatically intercepts the sensor raw data 8 seconds before the error occurs, generates incremental training samples through the sliding time window algorithm, and uses the stochastic gradient descent method with momentum term to update the fully connected layer parameters of the bidirectional long short-term memory network.

2. A laser engraving precision control system according to claim 1, characterized in that: In the model building and analysis unit (2), the feature preprocessing method of the time series feature mining model includes the following steps: The energy entropy feature sequence is decomposed into two channels using wavelet packets. The first channel extracts the energy distribution gradient vector in the 0.5-2kHz frequency band, and the second channel extracts the mutation pulse density in the 2-3kHz frequency band. The local linear embedding algorithm is used to perform nonlinear dimensionality reduction on the two-channel features to generate a 16-dimensional coupled feature vector. A residual connection module is added before the bidirectional LSTM input layer, and the current feature vector is weightedly fused with the previous five historical vectors and then input into the network.

3. The laser engraving precision control system according to claim 1, characterized in that: The attention mechanism optimization method of the model construction analysis unit (2) includes: A hybrid attention module is constructed to extract the spatial correlation weights of temporal features through a two-dimensional convolution kernel, generate a spatial attention map of feature nodes, apply temporal attention weights to the hidden layer output of the bidirectional long short-term memory network, perform Hadamard product operation with the spatial attention map after normalization using the hyperbolic tangent function, and set a dynamic activation factor to suppress the weights of pseudo-noise nodes.

4. The laser engraving precision control system according to claim 1, characterized in that: The network training method of the model building analysis unit (2) further includes: In the pre-training phase of the transfer learning framework, a benchmark model is constructed using three types of processing data: aluminum alloy, titanium alloy, and composite materials. In the online deployment phase, a domain adaptation mechanism is introduced to align the feature distribution differences under different working conditions through the gradient reversal layer, and a forget gate threshold is set. When the proportion of new working condition data exceeds the forget gate threshold, the network topology reconstruction function is enabled.

5. The laser engraving precision control system according to claim 1, characterized in that: The real-time detection method of the hardness of the processed material in the dynamic threshold adjustment unit (3) includes: A piezoelectric film sensor is installed on the rake face of the tool to collect the fluctuation signal of the three-dimensional component of the cutting force in real time. The Hilbert-Huang transform is performed on the fluctuation signal of the three-dimensional component to extract the instantaneous frequency and the slope of the cutting force envelope. A hardness inversion model is constructed based on the material plastic strain equation.

6. The laser engraving precision control system according to claim 1, characterized in that: The threshold offset calculation optimization method of the dynamic threshold adjustment unit (3) includes: A material hardness-thermomechanical coupling correction model is established, and a cutting temperature compensation term is introduced. When intermittent cutting is detected, the impact factor correction is automatically enabled in combination with the cutting temperature compensation term, and an exponential smoothing filter is applied to the final threshold.

7. The laser engraving precision control system according to claim 1, characterized in that: The dynamic threshold adjustment unit (3) also includes an adaptive adjustment strategy: A historical compensation decision database is constructed to calculate the optimal threshold offset values ​​under different combinations of processing material hardness and cutting temperature. A deep learning algorithm is used to dynamically optimize the parameters of the threshold offset calculation formula. When the processing accuracy fails to meet the standard after three consecutive compensations, the parameter regression test mode is triggered.

8. The laser engraving precision control system according to claim 1, characterized in that: The model building and analysis unit (2) also includes a timing prediction compensation mechanism: A Kalman filter is cascaded after the output layer of the bidirectional long short-term memory network to perform state estimation and correction on the compensation coefficient, and a compensation transfer function model is established. When the predicted compensation exceeds the physical limit, the constraint satisfaction algorithm is activated to replan the parameter trajectory.

9. The laser engraving precision control system according to claim 1, characterized in that: The dynamic threshold adjustment unit (3) further includes a multi-objective collaborative optimization module: A Pareto frontier analysis model is constructed to weigh the balance point between machining efficiency and tool life, a dynamic optimization function is defined, and then a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem in real time.

10. A method for implementing a laser engraving precision control system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Real-time data collection of processing equipment status is performed through vibration, current, and temperature sensors. Wavelet packet decomposition is used to extract energy entropy characteristics in the 0.5-3kHz frequency band to form a multi-dimensional time series feature matrix. S2: Input the 15-minute energy entropy sequence into the bidirectional LSTM model, combine residual connection and local linear embedding dimensionality reduction to generate 16-dimensional coupling features, optimize the weight distribution through the hybrid attention mechanism, and output the dynamic compensation coefficients of the spindle speed and feed rate; S3, based on the cutting force signal, invert the real-time hardness of the material and calculate the offset of the thermal-mechanical coupling correction threshold. When the compensation coefficient deviation exceeds the threshold, the CNC system is updated and an exponential smoothing filter is used to stabilize the threshold. S4: If the machining error exceeds the limit, the data before the error is intercepted for 8 seconds to generate incremental samples, the parameters of the LSTM fully connected layer are updated, the feature distributions of the new and old working conditions are aligned through the domain adaptation mechanism, and the network topology is reconstructed when the forget gate threshold is exceeded; S5. Utilize the filter to correct the compensation coefficient state, and adopt the Pareto frontier and genetic algorithm to dynamically optimize the balance point between machining efficiency and tool life.

Images