Intelligent optimization method for drill bit machining parameters based on machine learning

Abstract

The invention discloses a drill bit machining parameter intelligent optimization method based on machine learning, and particularly relates to the technical field of industrial intelligent manufacturing and machine learning, and the method comprises the steps: carrying out the real-time sensing and evaluation of a machining state through multi-modal sensing data fusion, achieving the dynamic precise prediction of a machining effect and a risk based on the online and offline dual-model cooperation, and achieving the real-time monitoring of a machining state. According to the real-time working condition and the production demand, the weight of the multi-objective function is dynamically adjusted and optimization is carried out to generate processing parameters, finally, the parameters are optimized, the search space of subsequent optimization is adaptively managed according to risk prediction, and an intelligent closed loop of perception, prediction, optimization and execution is formed. According to the method, comprehensive and accurate sensing of the machining state, continuous evolution of a prediction model, dynamic self-adaption of an optimization strategy and intelligence of closed-loop control are realized, so that the machining efficiency is improved, the service life of a cutter is prolonged, and the workpiece quality and the process safety are improved.

Description

Technical Field

[0001] This invention relates to the fields of industrial intelligent manufacturing and machine learning technology, and more specifically, to a method for intelligent optimization of drill bit processing parameters based on machine learning. Background Technology

[0002] In the field of machining, especially drilling, the selection of machining parameters directly affects machining efficiency, tool life, and final workpiece quality. Traditionally, machining parameters have relied mainly on process manual recommendations, operator experience, or the "trial and error" method. This approach is not only inefficient but also ill-suited to dynamic conditions such as tool wear and material batch fluctuations, failing to optimize machining performance. With the development of intelligent manufacturing technology, some sensor-based machining status detection methods have emerged. By extracting signal features and comparing them with preset thresholds, tool wear warnings or chatter identification can be achieved. Furthermore, some research attempts to establish empirical models or simple machine learning models between machining parameters and results for post-hoc parameter evaluation or static optimization within a limited range.

[0003] However, in practical use, it still has some shortcomings. For example, most methods rely on a single type of physical signal for analysis, which limits the dimensions of tool condition judgment and makes it impossible to effectively capture and quantify visual wear such as microscopic chipping of the cutting edge, resulting in incomplete condition assessment and insufficient early warning accuracy. The constructed prediction models are often static and lack online self-updating mechanisms, making it difficult to adapt to differences in the material properties of different batches of workpieces or the performance degradation of the tool itself. The model's generalization ability and long-term prediction accuracy will decrease over time. Existing parameter optimization methods mostly use objective functions with fixed weights, which cannot dynamically adjust the optimization direction according to real-time production needs and processing risks, resulting in rigid strategies. Most systems lack closed-loop adaptive control logic, and the search space after parameter optimization is fixed, making it impossible to achieve an intelligent balance between ensuring safety and exploring the efficient range, thus limiting the overall performance improvement potential of the system in long-term operation. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a machine learning-based intelligent optimization method for drill bit processing parameters, which addresses the problems mentioned in the background section through the following approach.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent optimization method for drill bit machining parameters based on machine learning, comprising:

[0006] S1: Simultaneously acquire dynamic physical signals and tool edge images during the machining process; perform time-frequency analysis on the dynamic physical signals to extract time-frequency domain features, and extract visual wear features from the tool edge images through a convolutional neural network; deeply fuse the two types of features to generate a comprehensive state representation vector, and determine the abnormal risk level in real time based on this vector;

[0007] S2: Input the comprehensive state representation vector and current machining parameters into the online real-time prediction model, and output the tool wear rate, workpiece surface quality change trend and quantified real-time risk prediction value in real time; in parallel, the offline deep analysis model is trained periodically, and its learning results are transferred to the online model through knowledge distillation to continuously improve the prediction accuracy;

[0008] S3: Establish an objective function that balances material removal rate, tool wear state prediction, and surface quality prediction; dynamically adjust the weights of each indicator in the objective function based on real-time order priority, workpiece batch quantity, and the aforementioned abnormal risk level; within process constraints, search for the optimal combination of spindle speed and feed rate parameters using the adjusted objective function;

[0009] S4: Execute the optimal combination of parameters; simultaneously, dynamically adjust the search space boundary in subsequent parameter optimization iterations based on the real-time risk prediction value: when the predicted risk is higher than the threshold, shrink the boundary to ensure processing safety; when the real-time risk prediction value is consistently lower than the safety threshold, widen the boundary to explore a more efficient processing parameter region.

[0010] The technical effects and advantages of this invention are as follows:

[0011] 1. Achieved comprehensive and accurate processing status perception and risk assessment: By synchronously acquiring dynamic physical signals such as vibration and acoustic emission and tool cutting edge images, and using time-frequency analysis and convolutional neural networks to extract features and then performing deep fusion, a comprehensive status representation vector with complementary information is generated; overcoming the limitations of a single signal source, it can not only perceive macroscopic anomalies such as chatter, but also accurately quantify microscopic visual wear such as cutting edge chipping, thereby improving the comprehensiveness of status assessment and the accuracy of anomaly warning;

[0012] 2. A continuously evolving high-precision prediction model has been constructed: A dual-model collaborative architecture of "real-time prediction by an online lightweight model + periodic training by an offline deep model" has been adopted. The learning results of the offline model are transferred to the online model through knowledge distillation technology. This enables the online prediction model to continuously update and optimize itself without interrupting production, effectively adapt to changes such as tool degradation and material fluctuations, and maintain high accuracy and high generalization ability in predicting tool wear rate, surface quality and overall risk in the long term.

[0013] 3. Achieved multi-objective adaptive optimization for dynamic needs: By establishing core objective functions for composite material removal rate, predicted wear state and predicted surface quality, and designing a dynamic weight allocator to dynamically adjust the weights of each indicator according to real-time order priority, remaining quantity of workpiece batch and abnormal risk level, the parameter optimization process can flexibly respond to actual production needs and changes in working conditions.

[0014] 4. A smart closed-loop control that balances safety and efficiency has been formed: The system not only outputs and executes the optimal parameters, but also dynamically manages the search space boundary of subsequent optimization iterations based on real-time risk prediction values; enabling the entire system to automatically find the best balance between the safety red line and the performance boundary during long-term operation, achieving robust and proactive adaptive control.

[0015] 5. Improved overall processing efficiency and economic benefits: The system integrates condition monitoring, trend prediction, parameter optimization and execution control into an intelligent closed-loop system, which reduces downtime caused by abnormal tool damage, extends tool life, stabilizes and improves workpiece processing quality, and maximizes material removal rate. This comprehensively improves the overall economic benefits and intelligent level of drilling processing from multiple dimensions such as cost reduction, efficiency improvement, quality improvement and safety assurance. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0017] Figure 2 This is a schematic diagram of the multimodal state sensing structure of the present invention.

[0018] Figure 3 This is a schematic diagram of the dual-model collaborative prediction structure of the present invention.

[0019] Figure 4 This is a schematic diagram of the dynamic optimization and execution closed-loop structure of the present invention.

[0020] Figure 5 This is a graph showing the relationship between the optimized objective function value and the objective function of this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] refer to Figures 1-5 The machine learning-based intelligent optimization method for drill bit machining parameters shown includes:

[0023] S1: Simultaneously acquire dynamic physical signals and tool edge images during the machining process; perform time-frequency analysis on the dynamic physical signals to extract time-frequency domain features, and extract visual wear features from the tool edge images through a convolutional neural network; deeply fuse the two types of features to generate a comprehensive state representation vector, and determine the abnormal risk level in real time based on this vector;

[0024] S2: Input the comprehensive state representation vector and current machining parameters into the online real-time prediction model, and output the tool wear rate, workpiece surface quality change trend and quantified real-time risk prediction value in real time; in parallel, the offline deep analysis model is trained periodically, and its learning results are transferred to the online model through knowledge distillation to continuously improve the prediction accuracy;

[0025] S3: Establish an objective function that balances material removal rate, tool wear state prediction, and surface quality prediction; dynamically adjust the weights of each indicator in the objective function based on real-time order priority, workpiece batch quantity, and the aforementioned abnormal risk level; within process constraints, search for the optimal combination of spindle speed and feed rate parameters using the adjusted objective function;

[0026] S4: Execute the optimal combination of parameters; simultaneously, dynamically adjust the search space boundary in subsequent parameter optimization iterations based on the real-time risk prediction value: when the predicted risk is higher than the threshold, shrink the boundary to ensure processing safety; when the real-time risk prediction value is consistently lower than the safety threshold, widen the boundary to explore a more efficient processing parameter region.

[0027] S1: Real-time perception, feature extraction, fusion, and preliminary risk assessment of multimodal processing status data. Specific implementation is as follows:

[0028] S101. Data Synchronous Acquisition: A three-axis high-frequency vibration acceleration sensor with a sampling frequency of not less than 20kHz installed on the machine tool spindle box and an acoustic emission sensor with a sampling frequency of not less than 1MHz installed near the workpiece fixture synchronously acquire dynamic physical signals during the machining process.

[0029] Meanwhile, the high-definition industrial camera integrated at the end of the coolant endoscope line can be configured with a frame rate of 10-30fps, triggering and capturing microscopic images of the tool cutting edge area at the moment when the drill bit periodically withdraws from the workpiece for chip removal.

[0030] By using a hardware trigger card or a software synchronization protocol based on the CNC system clock, the physical signal flow is ensured to be time-aligned with the image acquisition time, with the error controlled within milliseconds.

[0031] S102. Feature Extraction:

[0032] Dynamic physical signal processing: The acquired vibration and acoustic emission raw signals are preprocessed, including noise reduction and normalization; then, time-frequency analysis is performed on them, specifically using short-time Fourier transform or continuous wavelet transform, to extract a set of time-frequency domain features, including but not limited to: energy, power spectral entropy, and wavelet packet energy coefficients of a specific frequency band.

[0033] Visual image processing: The captured tool cutting edge image is preprocessed, including grayscale conversion, contrast enhancement, and region of interest (ROI) cropping, with the ROI focused on the main cutting edge and corners;

[0034] The preprocessed image is input into a pre-trained convolutional neural network for feature extraction. This CNN can be a lightweight network structure, and the output of the layer before the final fully connected layer is extracted as a high-dimensional visual wear feature vector, which can characterize visual morphological changes such as edge chipping and the width of the wear band on the back face.

[0035] S103. Feature Fusion and Risk Level Determination: The extracted time-frequency domain feature vector (dimension m) and visual wear feature vector (dimension n) are deeply fused. Specifically, a combination of feature concatenation and nonlinear mapping is used: first, the two vectors are concatenated into an m+n dimensional joint vector, and then this vector is input into a shallow fully connected neural network (containing 1-2 hidden layers). This fusion network, through supervised training, learns to map heterogeneous features to a unified, complementary, comprehensive state representation vector (dimension p, p...). <m+n)。

[0036] Based on this comprehensive state representation vector, an anomaly risk level is determined in real time through a pre-built risk assessment module based on rule-based logic or a lightweight classifier. Specifically:

[0037] The risk level can be discretized into three levels: "low", "medium" and "high";

[0038] The judgment rule can be as follows: if the feature values ​​representing high-frequency vibration energy and the degree of cutting edge damage in the vector simultaneously exceed a certain percentage of their historical operating baseline, it is judged as a "high" risk level; if only one exceeds, it is judged as a "medium" risk level; otherwise, it is a "low" risk level. This level serves as a rapid and qualitative judgment of the instantaneous abnormal state of the machining system.

[0039] It should be further explained that, in order to achieve a precise temporal correspondence between vibration, acoustic emission signals and visual images, the hardware trigger card or software synchronization protocol must ensure that the same high-precision timestamp is applied to the dynamic physical signal data stream at the same moment of exposure sampling by the image sensor. In post-processing, the signal segments within the corresponding time window are aligned and analyzed based on these timestamps, thereby ensuring that the subsequently extracted time-frequency domain features and visual features reflect the same processing transient.

[0040] In a preferred embodiment, the convolutional neural network is first initialized by transfer learning on a publicly available large image dataset, and then subjected to supervised fine-tuning training using a large number of labeled tool wear images. The tool wear image labeling information includes wear type, wear area, and wear level.

[0041] The input to the shallow fully connected neural network is a concatenated m+n dimensional vector. The supervision signals used in its training process are the actual tool wear and workpiece surface roughness values ​​obtained through offline precision measurement corresponding to that moment. The network learns by minimizing the regression error between its output (the comprehensive state representation vector) and these actual process states. Therefore, this vector is essentially a data-driven compressed representation strongly correlated with the final process quality indicators, rather than a simple feature stacking.

[0042] Further examples of the logic for determining “abnormal risk level”: The “historical operating baseline” usually refers to the collection of data and calculation of the statistical distribution of each characteristic value over a period of time when the initial tool is intact and the machining is stable; “exceeding a certain percentage” can be specifically quantified as: when the characteristic value exceeds its baseline mean plus k times the standard deviation (k=3), it is considered abnormal.

[0043] The "rule-based logic or lightweight classifier" can be specifically implemented as a decision tree or linear discriminant analysis model, with its input being a comprehensive state representation vector and its output being a discrete risk level. This module is trained using historical abnormal operating condition data to learn the feature boundaries that distinguish normal conditions from different levels of abnormal states; historical abnormal operating condition data includes data fragments of moments such as chatter occurrence or tool chipping.

[0044] S2: Implement processing effect and risk prediction based on dual-model collaboration, and achieve online self-updating of the model through knowledge distillation. The specific implementation is as follows:

[0045] S201. Construction and operation of online real-time prediction models:

[0046] Construct a lightweight online real-time prediction model, such as a streamlined fully connected neural network or a small recurrent neural network (RNN). The model's input layer receives a comprehensive state representation vector (dimension p) and current machining parameters, including spindle speed and feed rate, which can be normalized. The model has three output heads, corresponding to three regression tasks:

[0047] Output 1: Tool wear rate;

[0048] Output 2: The trend of workpiece surface quality change, which can be quantified as the predicted instantaneous value or change gradient of surface roughness (Ra);

[0049] Output 3: A quantified real-time risk prediction value, which is a continuous value between 0 and 1, comprehensively reflecting the probability of process instability, tool failure, or quality deviation based on the current state and trend.

[0050] The model is deployed on edge computing devices or industrial PCs with real-time computing capabilities, ensuring that the single inference latency is less than the 50-millisecond cycle required by the control system.

[0051] S202. Periodic training of offline deep analysis models:

[0052] An offline deep analysis model is run in parallel on a backend server or in the cloud. This model has a deeper network structure to achieve higher prediction accuracy. The model is trained using historical data. The training dataset includes all "comprehensive state representation vectors - machining parameters" collected in historical machining as input features, and the corresponding "actual tool wear," "measured surface roughness," and "risk labels" marked by experts, obtained through precise post-processing measurements, as supervision signals. The training process is automatically triggered once every 50 workpieces machined or every 8 hours.

[0053] S203. Knowledge Distillation-Based Model Update: After each offline model retraining, the knowledge distillation process is initiated. The specific implementation is as follows:

[0054] Knowledge Extraction: Prepare a transfer dataset, which can be a subset of recent data. Use a trained offline deep analytics model as a "teacher model" to infer from this dataset, outputting not only its predictions but, more importantly, its soft target, i.e., the output layer before Softmax activation. Value, these It contains rich inter-category relationships and uncertainty information learned by the teacher model.

[0055] Knowledge transfer: The input features of the above transfer dataset, namely the comprehensive state representation vector and processing parameters, are input into the online real-time prediction model as the "student model," and the soft objective generated by the teacher model is used. As part of the supervision signal, it is combined with the real label to form the loss function for fine-tuning the student model; the loss function typically used is:

[0056] ,in and These are the weighting coefficients.

[0057] Model replacement: After the fine-tuned training converges, the running online real-time prediction model is seamlessly replaced with the updated student model parameters, achieving continuous improvement in prediction accuracy. This process is completed using hot deployment technology to ensure uninterrupted control flow.

[0058] It needs to be further explained that, , ;

[0059] Ensuring the basic accuracy of the prediction results (primary objective): This ensures that the student model's output is directly aligned with the true measurement labels, which is fundamental to optimizing decision reliability. Weights are assigned to them. The aim is to prioritize ensuring that the student model after distillation does not deviate from the actual process rules, avoid introducing deviations due to excessive imitation of the teacher model, and ensure the stability of the basic performance of the optimization system;

[0060] Effective transfer of generalized knowledge from the teacher model (secondary but key objective): The term enables the student model to learn the inter-category relationships and uncertainties inherent in the teacher model's output, and assigns them weights. This allows student models to absorb smoother, more generalizable mapping patterns learned by teacher models from massive amounts of data, without excessively interfering with basic accuracy.

[0061] Further dynamic adjustment strategies: During the iterative process of distillation training, dynamic adjustment strategies can be adopted to optimize the results.

[0062] Initial stage: can be adopted , This allows student models to quickly approximate real labels and establish a basic level of accurate prediction capabilities.

[0063] Mid-term phase: Gradually adjust the weights to , At this point, the student model has achieved basic accuracy and begins to focus on absorbing the generalized knowledge from the teacher model.

[0064] Post-production fine-tuning stage: can be adjusted again to , This allows the student model to internalize the output distribution characteristics of the teacher model more deeply, further improving its performance on marginal cases.

[0065] S3: Based on real-time status and business requirements, the optimal combination of processing parameters is solved through a multi-objective optimization function with dynamically adjusted weights. The specific implementation is as follows:

[0066] S301. Establishment of the core optimization objective function: Construct the core optimization objective function in the following form. :

[0067]

[0068] in:

[0069] As a positive indicator of processing efficiency, it can be determined based on the spindle speed. Feed rate and drill bit diameter Through formula Calculate directly, or use its normalized value.

[0070] The cost function is used to predict the tool wear state. This state is based on the tool wear rate output by the S2 online model. The predicted instantaneous wear amount is obtained by integrating the current cumulative processing time or number of holes. .like, This indicates that wear costs increase non-linearly (squarely) with the amount of wear.

[0071] The cost function for predicting workpiece surface quality is based on surface roughness. Characterization, The value comes directly from the output of the S2 model. For example, This indicates that when the predicted roughness exceeds the technical requirement value... In such cases, punitive costs are incurred.

[0072] These are the dynamic weight coefficients of the three indicators, and they satisfy the normalization constraint. .

[0073] It should be further noted that some of the basic scene parameters in this experiment are as follows: As a positive indicator of processing efficiency, the calculation formula is: ;

[0074] The cost function for predicting tool wear conditions is calculated using the following formula: ;

[0075] The cost function for predicting workpiece surface quality is calculated using the following formula: ;

[0076] Dynamic weighting coefficients The table is as follows:

[0077]

[0078] The data in the table shows that: [Base weight] As the spindle speed and feed rate increase, the material removal rate, tool wear cost, and surface quality cost all increase, but the overall objective function value shows an upward trend.

[0079] S302. Implementation of Dynamic Weight Allocator: Weight Coefficients It is not fixed, but is adjusted by a dynamic weight allocator based on the following real-time inputs:

[0080] Order Priority : Obtained from the Manufacturing Execution System (MES), it can be quantified into three levels: 1 (low), 2 (medium), and 3 (high).

[0081] Remaining quantity of workpiece batch : From the production plan.

[0082] Abnormal risk level "Low" = 1, "Medium" = 2, "High" = 3.

[0083] The logic for weight adjustment is implemented through a pre-defined rules engine or a small configuration network. An example is shown below:

[0084] Rule engine example:

[0085] Base weight: set to .

[0086] Based on priority adjustment: If ,but It places more emphasis on efficiency; if but It places more emphasis on tool life.

[0087] Adjustment based on remaining quantity: If That is, the batch is about to be completed. It places greater emphasis on ensuring the quality of the last batch of workpieces.

[0088] Based on risk level adjustment: If ,but That is, reduce the efficiency weight to ensure safety, and increase accordingly. and Normalization is required after each adjustment.

[0089] Network configuration example: , , The input features are fed into a lightweight neural network, which directly outputs a normalized weight vector. The network can be trained through supervised learning using historical optimization decision data.

[0090] S303. Online Real-Time Parameter Optimization: Within each optimization cycle, the final objective function is determined using dynamically adjusted weighting coefficients. Within the constraints of machine tool safety processes, such as the upper limit of spindle speed... Upper limit of feed force Power limit These constraints can be preset based on machine tool manuals and material properties, and the optimal parameter combination can be searched using efficient real-time optimization algorithms. .

[0091] Optimization algorithm selection: Since the optimization cycle is usually in the second or sub-second range, and the main variables are rotation speed and feed, methods such as sequential quadratic programming, fast variants of particle swarm optimization, or single-step fast inference of Bayesian optimization can be used.

[0092] Search process: The optimization algorithm calls the online real-time prediction model trained in S2 as the objective function. and A fast evaluator for terms that iteratively evaluates different terms within the constraint boundaries. Combine the corresponding FF values ​​until the optimal solution that maximizes FF is found. .

[0093] S4: Completes the closed loop of optimization decision execution and realizes intelligent management of the optimizer's own behavior (search boundary), which is the ultimate manifestation of adaptive control logic. The specific implementation is as follows:

[0094] S401. Parameter Issuance and Execution: The optimal spindle speed obtained through optimization in S3 will be... With feed rate The parameter combination is sent to the machine tool's numerical control system (CNC) for execution via a preset communication interface. The sending process includes:

[0095] Instruction encapsulation and verification: Encapsulate parameter values ​​into G-code instructions or direct data packets that can be recognized by the CNC system, and perform range and format verification;

[0096] Smooth transition processing: To avoid the impact of sudden parameter changes on the machining process, ramp function or S-curve acceleration programming is used to smoothly transition the parameters from the current value to the target value over several control cycles. , ;

[0097] Execution confirmation: Monitor the actual spindle speed and feed rate fed back by the CNC system to confirm that the command has been received and executed.

[0098] S402. Dynamic Management of Search Boundary Based on Risk Prediction: To achieve long-term and safe optimization, the system maintains a dynamic parameter search space boundary, which defines the spindle speed range allowed for the subsequent optimization algorithm in S3. and feed rate range The initial values ​​of these boundaries are set as the absolute safety boundaries allowed by the machine tool and process.

[0099] The core of dynamic management is based on the real-time risk prediction values ​​output from S2. To adjust these edges, real-time risk prediction values It is a continuous quantity boundary between 0 and 1. Specific implementation includes:

[0100] Security threshold setting: Preset one or more security thresholds, such as setting a master threshold. Warning threshold These thresholds can be configured based on the historical performance of the workpiece material and tool type.

[0101] Boundary adjustment logic:

[0102] when When the system determines that it is currently in a high-risk state, and the primary goal is to ensure safety, it actively narrows the search boundary. For example, it raises the upper limit of the rotational speed. Temporarily set the feed rate to 90% of the current safe operating speed, thus limiting the feed rate. Set to 80% of the current value. This contraction is preventative and designed to limit the optimization algorithm from choosing overly aggressive parameters in the next cycle.

[0103] when Furthermore, if this state has persisted for three optimization cycles (predicting low and stable risk): the system determines the operating condition is very stable and has the potential to achieve higher efficiency. At this point, the search boundary should be appropriately relaxed, moving closer to the initial absolute safety boundary. For example, […]. and Each time, the range is increased by a step size of 1%, such as the absolute range, until the initial boundary is restored or the risk prediction value increases again. This allows the optimization algorithm to explore potential efficient parameter points in a wider space.

[0104] when (Medium prediction risk): Keep the current search boundaries unchanged and adopt a robust optimization strategy.

[0105] Implementation of boundary adjustment: The above logic is implemented through a dynamic constraint boundary management module. This module is triggered at the end of each optimization cycle, reading the latest... The new boundary value is calculated according to the above rules and updated to the optimization algorithm of S3 as the constraint condition for its next iteration.

[0106] It should be further explained that the main threshold Warning threshold ;

[0107] Main threshold The setting is based on the following: This threshold defines the critical point at which the system must take explicit protective intervention. Its primary goal is to absolutely prevent the processing process from entering an irreversible failure state. The value of 0.7 corresponds to a high level of confidence or probability of failure, which means that when the real-time risk prediction value of the online model exceeds this limit, the system judges that the processing process is highly likely to experience serious problems in the short term. At the same time, it avoids the system from triggering drastic intervention due to small fluctuations in the prediction value, thus ensuring the stability of the control.

[0108] Warning threshold The setting is based on the following: This threshold defines the conditions under which the system can safely try more efficient processing parameters. Its goal is to actively explore the performance boundary when the risk is extremely low. The value of 0.4 is lower than the main threshold, indicating that the system is in a very stable and reliable state. Setting the threshold at this level means that the system is only allowed to "slow down" when the risk prediction value is consistently lower than this line, that is, to relax the parameter search boundary.

[0109] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0110] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based intelligent optimization method for drill bit machining parameters, characterized in that, Including: S1: Synchronously acquire the dynamic physical signals and tool edge images during the machining process; perform time-frequency analysis on the dynamic physical signals to extract time-frequency domain features, and extract visual wear features from the tool edge images through a convolutional neural network; deeply fuse the two types of features to generate a comprehensive state representation vector, and accordingly determine the abnormal risk level in real time; S2: Input the comprehensive state representation vector and the current machining parameters into an online real-time prediction model to output the tool wear rate, the change trend of the workpiece surface quality, and the quantified real-time risk prediction value in real time; in parallel, periodically train an offline deep analysis model, and transfer its learning results to the online model through knowledge distillation to continuously improve the prediction accuracy; S3: Establish an objective function that weighs the material removal rate, predicts the tool wear state, and predicts the surface quality; dynamically adjust the weights of each index in the objective function according to the real-time order priority, the number of workpiece batches, and the abnormal risk level; within the process constraints, search for the optimal spindle speed and feed speed parameter combination with the adjusted objective function; S4: Issue and execute the optimal parameter combination; at the same time, dynamically adjust the search space boundary in the subsequent parameter optimization iteration according to the real-time risk prediction value: when the predicted risk is higher than the threshold, shrink the boundary to ensure machining safety; when the real-time risk prediction value is continuously lower than the safety threshold, relax the boundary to explore a higher-efficiency machining parameter region.

2. The intelligent optimization method for drill bit machining parameters based on machine learning according to claim 1, characterized in that, The comprehensive state representation vector includes: after splicing the time-frequency domain feature vector and the visual wear feature vector, input them into a shallow fully connected neural network for non-linear mapping and feature compression to generate a comprehensive state representation vector with a dimension of p and p < m + n, where m is the dimension of the time-frequency domain feature vector and n is the dimension of the visual wear feature vector.

3. The intelligent optimization method for drill bit machining parameters based on machine learning according to claim 1, characterized in that, The abnormal risk level includes: based on the comprehensive state representation vector, determine the discretized "low", "medium", and "high" three-level risk levels in real time through a preset rule-based logic or lightweight classifier.

4. The intelligent optimization method for drill bit machining parameters based on machine learning according to claim 1, characterized in that, The tool wear rate and the change trend of the workpiece surface quality include: the regression task results respectively output by three output heads of the online real-time prediction model, where the change trend of the workpiece surface quality is specifically quantified as the predicted instantaneous value of the surface roughness or its change gradient.

5. The intelligent optimization method for drill bit machining parameters based on machine learning according to claim 1, characterized in that, The real-time risk prediction value includes: input the comprehensive state representation vector and the current machining parameters into the online real-time prediction model, and perform regression calculation through the third output head of the model to output a continuous value between 0 and 1, which comprehensively reflects the probability of process instability, tool failure, or poor workpiece quality predicted based on the current state.

6. The intelligent optimization method for drill bit machining parameters based on machine learning according to claim 1, characterized in that, The prediction accuracy includes: build a dual-model architecture of "online real-time prediction model + offline deep analysis model", and the offline model completes a periodic training every 50 workpieces processed or every 8 hours using the historical comprehensive state representation vector, the corresponding machining parameters, and the corresponding measured process data; use the recent data as the transfer dataset, and the offline model outputs a soft target, and the online model fine-tunes in combination with the soft target and the true label; after the fine-tuning converges, replace the original model with the updated online model through hot deployment.

7. The intelligent optimization method for drill bit machining parameters based on machine learning according to claim 1, characterized in that, The objective function includes a multi-objective function constructed with material removal rate, predicted tool wear cost, and predicted working surface quality cost as optimization terms. Specifically, it is the weighted sum of the processing efficiency term minus the wear cost term and the surface quality cost term. The weight coefficients of each term are adjusted in real time according to the real-time order priority, the remaining quantity of workpiece batch, and the abnormal risk level through a dynamic weight allocator.

8. The intelligent optimization method for drill bit machining parameters based on machine learning according to claim 1, characterized in that, The optimal combination of spindle speed and feed rate parameters includes: within process constraints such as upper limits for spindle speed, feed force, and power, using a dynamically adjusted objective function. With the goal of maximizing, the optimal solution is searched by using real-time optimization algorithms such as sequential quadratic programming, fast particle swarm optimization, or Bayesian optimization, and by calling an online real-time prediction model as an evaluator.

9. The intelligent optimization method for drill bit machining parameters based on machine learning according to claim 1, characterized in that, The search space boundary includes: when At this time: In a high-risk state, actively shrink the search boundary, set the upper limit of rotational speed to 90% of the current safe rotational speed, and set the upper limit of feed rate to 80% of the current value; when If this state has persisted for 3 optimization cycles: widen the boundary by a step of 1% of the absolute range until it returns to the initial boundary or the risk prediction value increases again; when Time (medium risk of prediction): Keep the current search boundaries unchanged.

Images