Method for Predicting and Optimizing the Quality of Ultrasonic Vibration-Assisted Drilling Using Carbon Fiber Composite Materials

Abstract

This invention discloses a method for predicting drilling quality and optimizing parameters using ultrasonic vibration-assisted drilling of carbon fiber composite materials. The method constructs a multi-source sensing system to collect drilling process parameters, tool status information, and dynamic signals from the drilling process. It establishes a rapid parameterized digital twin model, outputting physical prior information such as drilling load and layered risk. Based on digital twin samples and actual drilling samples, a static-dynamic hybrid sample library is constructed, and a physical constraint temporal fusion network, including static parameter branches, dynamic temporal branches, and feature fusion modules, is trained to predict drilling quality indicators and their prediction uncertainties. Furthermore, active learning and incremental model updates are performed based on the prediction uncertainty and layered risk information. Constrained multi-objective optimization is then performed based on the trained model to output a set of optimal process parameters that satisfy both processing efficiency and quality constraints. This invention improves the accuracy of online drilling quality prediction, the model's adaptability across working conditions, and the efficiency of process parameter optimization.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic vibration-assisted drilling technology for carbon fiber composite materials, specifically a method for predicting the quality and optimizing the parameters of ultrasonic vibration-assisted drilling for carbon fiber composite materials. Background Technology

[0002] Carbon fiber composites, due to their high specific strength, high specific modulus, corrosion resistance, and fatigue resistance, are widely used in aerospace, high-end equipment, and rail transportation. These composite components typically require extensive hole drilling during assembly, and the quality of this drilling directly affects connection strength, assembly accuracy, and service reliability. Ultrasonic vibration-assisted drilling, by superimposing high-frequency micro-amplitude vibrations during drilling, can reduce cutting loads and improve hole quality to some extent, thus gradually becoming an important process for high-quality hole drilling of carbon fiber composites.

[0003] However, carbon fiber composites are characterized by anisotropy, low interlaminar bonding strength, poor thermal conductivity, and significant brittle fracture. During ultrasonic vibration-assisted drilling, they are still prone to quality problems such as exit delamination, increased hole wall roughness, burrs, and accelerated tool wear. Existing drilling quality prediction methods mostly rely on offline simulation, single process parameter modeling, or purely data-driven models. These methods typically struggle to simultaneously utilize process parameters, tool condition, and dynamic signals from the drilling process to characterize the machining state online, resulting in insufficient stability of prediction results when tool wear changes, material batch fluctuations, or operating condition changes occur.

[0004] Furthermore, existing methods often only provide single-point predictions of borehole quality, lacking a quantitative expression of prediction reliability and making it difficult to determine whether the current prediction is within the model's credibility range. Simultaneously, quality prediction and process parameter optimization are usually disconnected; the prediction model cannot proactively supplement samples and update based on uncertain operating conditions, nor can it automatically output optimal process parameters while meeting processing efficiency and quality constraints. Therefore, there is an urgent need for a method for ultrasonic vibration-assisted borehole quality prediction and parameter self-optimization using carbon fiber composite materials. This method can integrate multi-source sensing information, introduce physical prior constraints, output prediction uncertainty, and form a closed loop of "online prediction, proactive updating, and parameter optimization." Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the quality and optimizing the parameters of ultrasonic vibration-assisted drilling of carbon fiber composite materials, in order to solve the problems of insufficient linearity, poor adaptability to tool wear and material fluctuations, lack of uncertainty output, and inability to optimize process parameters in a closed loop in existing methods.

[0006] To achieve the above objectives, the present invention proposes the following technical solution: a method for predicting and optimizing the quality of ultrasonic vibration-assisted drilling using carbon fiber composite materials, comprising the following steps:

[0007] Construct a multi-source sensing system for ultrasonic vibration-assisted drilling process to collect drilling process parameters, tool status information and dynamic signals of the drilling process.

[0008] A rapid parametric digital twin model corresponding to the ultrasonic vibration-assisted drilling process is established. The rapid parametric digital twin model is used to characterize the effects of ultrasonic vibration, feed motion, thermal softening effect and tool wear on drilling load and exit delamination risk, and outputs the physical prior information of the drilling process.

[0009] Based on parametric experimental design, the aforementioned physical prior information, and actual drilling measurement data, a static and dynamic hybrid sample library is constructed, and the deviation between the digital twin sample and the actual drilling sample is corrected.

[0010] A physical constraint temporal fusion network is constructed, which includes a static parameter branch for processing static input features, a dynamic temporal branch for processing dynamic process signals, and a feature fusion module for fusing static features and dynamic temporal features.

[0011] The physical constraint temporal fusion network is trained using the static-dynamic hybrid sample library to output the prediction results of borehole quality indicators and the prediction uncertainty corresponding to the prediction results.

[0012] Based on the predicted uncertainty and the hierarchical risk information output by the fast parameterized digital twin model, active learning sampling and incremental model updates are performed.

[0013] Based on the trained physical constraint temporal fusion network, a constrained multi-objective optimization model is established to optimize the process parameters of ultrasonic vibration assisted drilling and output the optimal set of process parameters that meet the constraints of processing efficiency and quality.

[0014] Furthermore, in this invention, the drilling process parameters collected by the multi-source sensing system include at least one of spindle speed, feed per revolution, ultrasonic frequency, and ultrasonic amplitude; the tool status information includes at least one of tool geometry parameters, tool coating type, and tool wear; and the dynamic process signal includes at least one of spindle current signal, acoustic emission signal, vibration signal, and temperature signal.

[0015] Furthermore, in this invention, the preprocessing of the dynamic process signal includes:

[0016] Establish a unified time reference using machine tool control signals or spindle encoder signals;

[0017] Timing alignment of dynamic process signals with different sampling rates;

[0018] Smoothing interpolation processing is performed on low-frequency dynamic process signals;

[0019] Extract envelope features, energy features, peak features, or statistical features from high-frequency dynamic process signals;

[0020] The aligned dynamic process signal is segmented according to a sliding time window and then normalized to obtain multi-channel timing data for inputting the dynamic timing branch.

[0021] Furthermore, in this invention, the rapidly parameterized digital twin model includes:

[0022] An ultrasonic kinematic model used to characterize the axial ultrasonic vibration displacement of a drill bit;

[0023] A chip thickness model used to determine the instantaneous undeformed chip thickness based on the coupling relationship between feed motion and ultrasonic vibration;

[0024] An intermittent contact model used to characterize the high-frequency contact and separation states between the tool and the workpiece;

[0025] An axial force prior model is used to integrate the instantaneous undeformed chip thickness, intermittent contact state, thermal softening effect and tool wear effect to obtain the axial force prior time sequence.

[0026] A stratification risk assessment model for determining export stratification risk based on axial force characteristics, material fracture properties, and laminate structural parameters during the export stage.

[0027] Furthermore, in this invention, the hierarchical risk assessment model determines the hierarchical risk index through the axial force impulse at the exit stage, the critical hierarchical thrust at the exit, the effect of ultrasonic load reduction, the effect of thermal softening, and the effect of tool wear. The hierarchical risk index is used as a physical consistency constraint in the training process of the physical constraint temporal fusion network, and as a risk constraint index in the active learning sampling and constrained multi-objective optimization process.

[0028] Furthermore, in this invention, the construction of the static-dynamic hybrid sample library includes:

[0029] The initial process parameter combination is generated using a parameter space sampling method;

[0030] The initial process parameter combination is simulated and calculated using the rapid parameterized digital twin model to obtain digital twin samples and their physical prior information;

[0031] Collect a small number of actual drill samples and obtain the actual drill quality measurement labels corresponding to the actual drill samples;

[0032] Based on the residual between the digital twin output and the actual drilling measurement results, a residual correction module is established to align the digital twin samples and the actual drilling samples in the virtual and real domains, thereby obtaining the static and dynamic hybrid sample library.

[0033] Furthermore, in this invention, in the physically constrained temporal fusion network:

[0034] The static parameter branch is used to extract a global low-dimensional representation of process parameters, material layup information, and tool status information;

[0035] The dynamic timing branch is used to extract local timing features and long-range dependence features of spindle current, acoustic emission, vibration and temperature;

[0036] The feature fusion module employs a cross-attention mechanism to modulate the temporal features output by the dynamic temporal branch with the global low-dimensional representation output by the static parameter branch.

[0037] The physical constraint temporal fusion network also includes a multi-task output layer, which is used to simultaneously output the predicted mean and prediction uncertainty of at least two of the following drilling quality indicators: peak axial force, layering factor, borehole wall roughness, and burr height.

[0038] Furthermore, in this invention, a joint training objective is used when training the physical constraint temporal fusion network, and the joint training objective includes:

[0039] Multi-task prediction loss for fitting borehole quality measured labels;

[0040] Physical consistency loss is used to constrain the relationship between prediction quality indicators and hierarchical risk information to maintain a physically monotonically consistent relationship.

[0041] Domain alignment loss used to reduce the difference in feature distribution between digital twins and real diamond samples;

[0042] Uncertainty calibration loss is used to constrain the prediction uncertainty to match the actual prediction error.

[0043] Furthermore, in this invention, the active learning imputation and incremental model update include:

[0044] An active learning evaluation index is constructed based on the maximum prediction uncertainty, sample novelty, and stratified risk information;

[0045] When the active learning evaluation index of the candidate sample meets the preset triggering conditions, the corresponding process parameter combination will be included in the actual drilling and sample replenishment list.

[0046] The quality of the drill samples obtained from the supplementary drilling was measured and features were extracted, and then added to the incremental cache pool;

[0047] When the number of samples in the incremental cache pool reaches the preset number, the new samples in the incremental cache pool are mixed with the historical samples in the experience replay buffer, and local network fine-tuning is performed.

[0048] The local network fine-tuning includes freezing part of the basic feature extraction layer and updating the network parameters of the feature fusion module and the multi-task output layer.

[0049] Furthermore, in this invention, the constrained multi-objective optimization model takes minimizing at least two of the following indicators: peak axial force, layering factor, hole wall roughness, and burr height, as its optimization objective, and uses at least one of the following constraints: lower limit of material removal rate, upper limit of layering risk, and upper limit of prediction uncertainty. The constrained multi-objective optimization model uses Bayesian optimization to search within the candidate process parameter space, and adjusts the acquisition function based on the probability that the candidate process parameters satisfy all constraints, outputting a Pareto process parameter set that satisfies the processing efficiency constraint and achieves optimal drilling quality.

[0050] Beneficial effects: The technical solution of this application has the following technical effects:

[0051] This invention constructs a multi-source sensing system to collect drilling process parameters, tool status information, and dynamic signals of the drilling process. Combined with the physical prior information output by a rapid parameterized digital twin model, it enables a more comprehensive online characterization of the machining state during ultrasonic vibration-assisted drilling. Compared to methods relying solely on static process parameters or offline simulation, this invention can simultaneously reflect the impact of process settings, tool wear, and dynamic machining response on drilling quality, thereby improving the accuracy of predicting quality indicators such as peak axial force, delamination factor, hole wall roughness, and burr height, as well as its adaptability across different working conditions.

[0052] This invention further constructs a physically constrained temporal fusion network, fusing static parameter features with dynamic temporal features, and utilizes physical prior information provided by a digital twin model to constrain the network training process. This ensures that the prediction results not only fit actual drilling data but also remain consistent with the physical laws of drilling. This reduces the risk of anomalous predictions from purely data-driven models under conditions of small samples, tool wear variations, or material batch fluctuations, thus improving the stability and reliability of the model in long-term applications in actual industrial settings.

[0053] Furthermore, this invention outputs the prediction uncertainty along with the borehole quality prediction result, and performs active learning and incremental updates based on the prediction uncertainty and hierarchical risk information, enabling the model to continuously adaptively correct itself for high-risk or low-confidence conditions. Simultaneously, a constrained multi-objective optimization model is established based on the trained model, outputting an optimal set of process parameters while satisfying processing efficiency and quality constraints. Therefore, this invention achieves closed-loop control from online borehole quality prediction to parameter self-optimization, reducing the number of blind experiments and improving the development efficiency and processing quality consistency of ultrasonic vibration-assisted drilling technology for carbon fiber composites.

[0054] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below can be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other.

[0055] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0056] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings, wherein:

[0057] Figure 1 This is a flowchart of the overall method of the present invention;

[0058] Figure 2 A schematic diagram of a multi-source sensing and data acquisition system and its data processing flow;

[0059] Figure 3 A schematic diagram illustrating the process of quickly parameterizing digital twin models;

[0060] Figure 4 This is a schematic diagram of a physically constrained temporal fusion network structure.

[0061] Figure 5 A flowchart for active learning and constrained multi-objective Bayesian optimization closed-loop process;

[0062] Figure 6 This is a diagram illustrating the principles and comparison of physical constraints / digital twin models. Detailed Implementation

[0063] The following description, in conjunction with the accompanying drawings and the online quality prediction and parameter self-optimization process of ultrasonic vibration-assisted drilling of carbon fiber composite materials, further illustrates the present invention. This embodiment is used to illustrate specific implementations of the present invention and does not constitute a limitation on the scope of protection; all equivalent substitutions, parameter adjustments, or structural modifications made based on the concept of the present invention should fall within the scope of protection of the present invention.

[0064] This embodiment uses a 2 mm thick M21 / T700 carbon fiber composite laminate as the machining object, and conducts a hole-making experiment using an 8 mm diameter ultrasonic vibration-assisted drilling system. The process parameters are set as follows: spindle speed 3000 to 12000 rpm, feed per revolution 0.01 to 0.08 mm / revolution, ultrasonic frequency 20 to 35 kHz, and ultrasonic amplitude 1 to 6 micrometers. Tool conditions include at least the tool tip angle, helix angle, coating type, and flank wear width.

[0065] Appendix Figure 1 The overall flow of the method of this invention is illustrated. Taking the ultrasonic vibration-assisted drilling process of carbon fiber composite materials as an example, the flow sequentially includes multi-source sensing data acquisition, rapid parametric digital twin modeling, construction of a static-dynamic hybrid sample library, training of a physically constrained temporal fusion network, online prediction of drilling quality, active learning incremental updates, and constrained multi-objective Bayesian optimization. Through this flow, this invention integrates process state perception, physical prior modeling, data-driven prediction, and process parameter optimization into a closed-loop system, enabling the prediction results to guide the selection of process parameters.

[0066] Appendix Figure 2 The multi-source sensing and acquisition system and data processing flow are illustrated. This system is used to synchronously acquire static information such as spindle speed, feed per revolution, ultrasonic frequency, amplitude, tool geometry parameters, and tool wear, as well as dynamic process signals such as spindle current, acoustic emission, vibration, and temperature. After being aligned to a unified time base, matched sampling rates, segmented by a sliding window, and normalized, different types of signals are processed to form static parameter features and dynamic temporal features that can be used as input for subsequent network models, thus providing a complete data foundation for online borehole quality prediction.

[0067] Appendix Figure 3 The construction principle of a rapid parametric digital twin model is illustrated. This model calculates the instantaneous undeformed chip thickness based on ultrasonic vibration displacement, feed motion, and tool contact state, and obtains the a priori timing sequence of axial force by combining intermittent contact effect, thermal softening effect, and tool wear effect. Furthermore, it constructs a layered risk index based on the axial force impulse at the exit stage, critical layered thrust, and the effect of ultrasonic unloading. The role of this digital twin model is to provide the neural network with a priori constraints that conform to the physical laws of drilling, avoiding predictions that violate machining mechanisms due to the model relying solely on data fitting.

[0068] Appendix Figure 4 The structure of the physically constrained temporal fusion network is shown. This network includes a static parameter branch, a dynamic temporal branch, a cross-attention fusion layer, and a multi-task heteroscedasticity output layer. The static parameter branch extracts global representations of process parameters, tool status, and material layup information. The dynamic temporal branch extracts local temporal features and long-range dependencies from dynamic signals such as current, acoustic emission, vibration, and temperature. The cross-attention fusion layer implements weighted modulation of dynamic process features based on static condition information. The multi-task output layer simultaneously outputs quality indicators such as peak axial force, delamination factor, hole wall roughness, and burr height, along with their corresponding prediction uncertainties.

[0069] Appendix Figure 5 This paper illustrates a closed-loop process of active learning and constrained multi-objective Bayesian optimization. During online prediction, the system determines whether additional drilling samples are needed for the current working condition based on prediction uncertainty, sample novelty, and stratification risk index. When trigger conditions are met, the supplementary samples are added to the incremental cache pool, and the model is locally updated using an experience replay mechanism. In the process optimization stage, the system aims to minimize drilling quality indicators, constrained by processing efficiency, stratification risk, and prediction uncertainty. It outputs a Pareto-optimal set of process parameters through constrained multi-objective Bayesian optimization. This closed-loop process enables the model to continuously adjust with tool wear and material batch variations, improving the engineering reliability of the parameter recommendations.

[0070] Figure 6 The diagram consists of two parts. The left side is a flowchart of the "Physical Constraints / Digital Twin Model Principle," illustrating the working process of a rapidly parameterized digital twin model: taking process parameters (spindle speed, feed rate, ultrasonic parameters) and tool / material information (geometric parameters, material properties) as input, a rapidly parameterized digital twin model based on physical formulas and drilling mechanics (containing a cutting force calculation model and a tribological model) outputs a smooth, noise-free theoretical axial force prior time series reflecting the ideal state. The right side is a schematic diagram of "Comparison and Correction Requirements of Axial Force Time Series in Virtual and Real Domains," comparing the smooth prior time series curve calculated by the digital twin model with the sensor's measured signal (a sawtooth curve containing noise and dynamic characteristics) in the same coordinate system. The gray area indicates the deviation between the virtual and real domains, and points out that this deviation needs to be eliminated through residual correction. The conclusion at the bottom of the diagram highlights the core meaning of this diagram: it is necessary to integrate physical priors and measured data through the virtual-real domain alignment and residual correction module to improve model accuracy.

[0071] I. Establish a multi-source sensing system and acquire static parameters and dynamic process signals.

[0072] The purpose of this step is to obtain information input that reflects the entire process of "process setting - tool status - machining response," providing a data foundation for subsequent digital twin modeling, temporal fusion network training, and online prediction. Compared to solutions that only collect static process parameters such as spindle speed and feed rate, this step can simultaneously introduce dynamic process signals such as spindle current, acoustic emission, vibration, and temperature, enabling the model to identify tool wear, material batch fluctuations, and transient damage states.

[0073] In practical implementation, current sensors, acoustic emission sensors, triaxial vibration sensors, and infrared temperature measuring devices are arranged near the machine tool spindle and workpiece fixture. The acoustic emission signal sampling rate is preferably no less than 1MHz, the triaxial vibration signal sampling rate is preferably no less than 20kHz, and the response time of the infrared temperature measuring device is preferably no more than 10ms. The spindle feed enable signal issued by the CNC system is used as the global trigger source, and the spindle encoder pulse is used as the spatial period reference, uniformly mapping the signals from all sensors to the same drilling process and the same rotational period.

[0074] For signals with different sampling rates, this embodiment employs a differentiated alignment strategy: for low-frequency continuous signals such as temperature and spindle current, cubic spline interpolation is used for smooth upsampling; for acoustic emission and high-frequency vibration signals, the Hilbert envelope is extracted first, followed by sliding window root mean square, peak value, and energy feature resampling. After alignment, the dynamic signal is divided into a multi-channel time-series matrix according to a sliding time window related to the spindle rotation period, and Z-score normalization is performed.

[0075] The technical effects achieved in this step are as follows: dynamic signals are uniformly expressed in terms of time base and physical period, and data from different sensors can jointly describe the true state of the same drilling process; at the same time, high-frequency damage information is not lost due to direct downsampling, and low-frequency thermal load information is not subject to step distortion due to coarse upsampling. The underlying mechanism is that drilling damage in carbon fiber composite materials manifests not only as changes in macroscopic cutting load, but also as high-frequency energy abrupt changes caused by fiber breakage, interface debonding, and tool collision. Only by fusing multi-source heterogeneous signals on a unified physical time scale can an interpretable and learnable state representation be provided for subsequent models.

[0076] II. Establishing a rapid parameterized digital twin model

[0077] The purpose of this step is to provide the deep learning network with physical priors that conform to the mechanism of ultrasonic vibration-assisted drilling, avoiding the black box effect and anomalous predictions caused by the model relying entirely on a small amount of actual drilling data. The fast parameterized digital twin model does not pursue time-consuming full-scale finite element solutions, but instead uses parameterized expressions to quickly calculate the axial force prior timing, exit stage impulse, exit critical layer thrust, and layer risk index.

[0078] First, an axial ultrasonic vibration displacement model of the drill bit is established to describe the high-frequency micro-amplitude motion superimposed along the axial direction of the drill bit.

[0079] Formula (1): Axial ultrasonic vibration displacement model of drill bit

[0080]

[0081] In the formula, Indicates time axial ultrasonic displacement of the drill bit;

[0082] Indicates ultrasonic amplitude;

[0083] Indicates the ultrasonic frequency;

[0084] Indicates time;

[0085] Indicates the initial phase.

[0086] Then, the instantaneous undeformed chip thickness is calculated based on the coupling relationship between the conventional feed motion and the ultrasonic vibration displacement.

[0087] Formula (2): Instantaneous undeformed chip thickness model

[0088]

[0089] In the formula, Indicates time The instantaneous undeformed chip thickness;

[0090] This represents the chip thickness component generated by the normal feed motion;

[0091] This represents the axial ultrasonic displacement at the current moment;

[0092] This represents the axial ultrasonic displacement when adjacent cutting edges pass through the same point.

[0093] This indicates the time interval between adjacent cutting edges passing through the same point;

[0094] This indicates that the result within the parentheses is less than zero, and is used to characterize that no effective cutting occurs when the tool leaves the workpiece.

[0095] We further define the intermittent contact coefficient to characterize the periodic contact and separation phenomenon between the "tool and workpiece" caused by ultrasonic vibration.

[0096] Formula (3): Intermittent contact coefficient

[0097]

[0098] In the formula, Indicates time Intermittent contact coefficient;

[0099] when hour, This indicates that the tool and the workpiece are in effective contact and cutting state;

[0100] when hour, This indicates that the tool and workpiece are separated, and the axial cutting force is approximately zero.

[0101] After obtaining the instantaneous chip thickness and intermittent contact state, an a priori model of axial force is established. This model simultaneously considers the effects of chip thickness, ultrasonic intermittent contact, thermal softening effect, and tool wear effect on axial force.

[0102] Formula (4): Axial force prior model

[0103]

[0104] In the formula, Indicates time The a priori value of the axial force;

[0105] Indicates the intermittent contact coefficient;

[0106] Indicates the cutting force coefficient;

[0107] Indicates the instantaneous thickness of the undeformed chip;

[0108] Indicates the chip thickness index;

[0109] Indicates the thermal softening correction factor;

[0110] This indicates the tool wear correction factor.

[0111] Formula (5): Thermal softening correction factor

[0112]

[0113] In the formula, Indicates temperature The thermal softening correction factor is as follows;

[0114] Indicates the influence coefficient of material thermal softening;

[0115] Indicates the instantaneous temperature of the cutting zone;

[0116] Indicates ambient room temperature;

[0117] This indicates the glass transition temperature of the resin matrix.

[0118] Formula (6): Tool wear correction factor

[0119]

[0120] In the formula, Indicates the tool wear correction factor;

[0121] Indicates the wear effect coefficient;

[0122] Indicates the current wear width of the tool's flank face;

[0123] This indicates the set tool wear failure threshold;

[0124] This represents the nonlinear index of wear.

[0125] During the exit stage, the axial force has a direct driving effect on interlayer delamination. Therefore, this embodiment calculates the axial force impulse and the critical delamination thrust during the exit stage, and constructs a delamination risk index accordingly.

[0126] Formula (7): Axial force impulse at the exit stage

[0127]

[0128] In the formula, Indicates the axial force impulse during the export phase;

[0129] This indicates the starting moment when the drill bit enters the exit-sensitive phase;

[0130] This indicates the point at which the drill bit leaves the exit-sensitive stage.

[0131] This indicates the a priori timing of axial force.

[0132] Formula (8): Critical stratification thrust at the exit

[0133]

[0134] In the formula, Indicates the critical stratification thrust at the outlet;

[0135] Indicates the structural correction factor;

[0136] Indicates the interlaminar type I fracture toughness;

[0137] Represents the equivalent elastic modulus;

[0138] Indicates the thickness of the laminate.

[0139] Formula (9): Ultrasonic load reduction correction factor

[0140]

[0141] In the formula, This represents the ultrasonic load reduction correction factor;

[0142] Indicates a single ultrasonic vibration cycle;

[0143] Indicates the intermittent contact coefficient;

[0144] The integral term represents the cumulative effective contact time between the tool and the workpiece within a single ultrasonic cycle.

[0145] Formula (10): Stratified Risk Index

[0146]

[0147] In the formula, This indicates a tiered risk index;

[0148] Indicates the axial force impulse during the export phase;

[0149] Indicates the critical stratification thrust at the outlet;

[0150] Indicates the thermal softening correction factor;

[0151] Indicates the tool wear correction factor;

[0152] This represents the ultrasonic load reduction correction factor.

[0153] The technical benefits achieved in this step are as follows: the model can obtain physical priors of drilling load and layering risk at low computational cost, providing an interpretable basis for subsequent network training and constraint optimization. The underlying mechanism is that ultrasonic vibration-assisted drilling is not continuous cutting, but rather pulsed cutting accompanied by high-frequency contact-separation; the intermittent contact coefficient can identify the "air cutting" stage, avoiding the overestimation of axial force and exit impulse in traditional continuous cutting assumptions. The thermal softening correction factor reflects the change in cutting resistance caused by the thermal softening of the resin matrix, and the wear correction factor reflects the increase in friction and axial thrust caused by tool flank wear. These mechanisms together ensure consistency between the physical priors and the actual machining process.

[0154] III. Constructing a static-dynamic hybrid sample library and completing virtual-real domain alignment

[0155] The purpose of this step is to address the issues of expensive and insufficient sample size of carbon fiber composite drilling samples, as well as the difficulty in training purely data-driven models. In practice, Latin hypercube sampling or other space-filling sampling methods are used to generate initial parameter combinations within the range of parameters such as spindle speed, feed per revolution, ultrasonic frequency, amplitude, and tool wear. For each set of parameters, a fast parameterized digital twin model is called to obtain the axial force prior timing, exit stage impulse, and stratified risk index, thus forming digital twin samples.

[0156] Subsequently, a live drilling experiment was performed with a limited set of selected parameters to measure quality indicators such as peak axial force, exit delamination factor, borehole wall roughness, and burr height, while simultaneously acquiring multi-source dynamic process signals. The difference between the live drilling measurements and the digital twin predictions was used as a supervisory signal to train the residual correction module, enabling the model to specifically learn complex residuals that are difficult for the digital twin model to express, such as microscopic nonlinear friction, tool coating peeling, and the randomness of local fiber fracture.

[0157] The technical effect achieved in this step is that by using a digital twin to provide the physical framework and real-drill samples to correct nonlinear residuals, the model's prediction accuracy is improved while reducing the number of real-drill samples. The underlying mechanism is that the digital twin model can stably express the laws of ultrasonic kinematics, macroscopic mechanics, and fracture mechanics, while a small amount of real-drill data can compensate for the difficult-to-analyze random factors and local complex effects in the physical model. With the combination of these two, the network does not need to learn all cutting laws from scratch, thus converging more easily under small sample conditions and exhibiting better generalization ability across different working conditions.

[0158] IV. Constructing a Physically Constrained Temporal Fusion Network

[0159] The purpose of this step is to simultaneously process static process parameters and dynamic process signals, and establish a physical causal relationship between the two types of information. The physical constraint temporal fusion network includes a static parameter branch, a dynamic temporal branch, a cross-attention fusion layer, and a multi-task heteroscedastic output layer.

[0160] The static parameter branch receives inputs such as spindle speed, feed per revolution, ultrasonic frequency, amplitude, tool geometry parameters, tool coating type, lamination method, and tool wear. It extracts a global low-dimensional representation through a multi-layer fully connected network. This branch describes the "machining conditions and tool state." The dynamic temporal branch receives multi-channel time series data such as spindle current, acoustic emission, vibration, and temperature. It preferably employs a cascaded structure of a temporal convolutional network and a Transformer; the temporal convolutional network extracts local multi-scale impact features, while the Transformer extracts long-range dependency features.

[0161] The cross-attention fusion layer is used to explicitly express the modulation relationship between static parameters and dynamic signals. That is, the static branch output is used as the query vector, and the dynamic branch output is used as the key vector and value vector, so as to adaptively adjust the weights of different dynamic features according to process conditions and tool status.

[0162] Formula (11): Cross-attention fusion expression

[0163]

[0164] In the formula, This represents the features after cross-attention fusion;

[0165] This represents the query vector obtained by projecting the output of the static parameter branch;

[0166] This represents the key vector obtained by projecting the output of the dynamic timing branch;

[0167] This represents the value vector obtained by projecting the output of the dynamic timing branch;

[0168] Indicates the dimension of the key vector;

[0169] This represents the function for calculating normalized weights.

[0170] The multi-task heteroscedasticity output layer includes multiple parallel task heads, which output the predicted mean values ​​of peak axial force, delamination factor, hole wall roughness, and burr height, respectively, and simultaneously output the corresponding predicted variance. The predicted variance is used to characterize the confidence level of the current prediction result.

[0171] The technical effect achieved in this step is that the network can automatically select more critical dynamic process features based on different process conditions and tool wear states, improving the accuracy and stability of online prediction of drilling quality. The underlying mechanism is that the machining process is not determined solely by dynamic signals, but rather by static conditions such as spindle speed, feed rate, ultrasonic amplitude, and tool wear, which determine the physical meaning of the dynamic signals. For example, when tool wear intensifies, high-frequency distortion in acoustic emission and vibration becomes more effective in characterizing damage risk; when temperature increases, the explanatory power of temperature gradients on resin softening and hole wall quality is enhanced. The cross-attention mechanism embeds this physical causal modulation relationship into the network structure.

[0172] V. Training the network using a joint objective function

[0173] The purpose of this step is to enable the network to simultaneously possess data fitting ability, physical consistency, virtual-real domain adaptation ability, and uncertainty calibration ability. During training, a joint objective function is used, incorporating multi-task prediction loss, physical consistency loss, domain alignment loss, and uncertainty calibration loss into the optimization process.

[0174] Formula (12): Joint objective function

[0175]

[0176] In the formula, Describe the joint objective function;

[0177] This represents the multi-task heteroscedasticity prediction loss;

[0178] Indicates the loss of physical consistency;

[0179] Indicates domain alignment loss;

[0180] Indicates the uncertainty calibration loss;

[0181] , , , These represent the weighting coefficients of each loss term.

[0182] Formula (13): Multi-task heteroscedasticity prediction loss

[0183]

[0184] In the formula, This represents the multi-task heteroscedasticity prediction loss;

[0185] Indicates the number of predicted tasks;

[0186] Indicates the first The actual measured values ​​of each quality indicator;

[0187] Indicates the first The predicted mean of each quality indicator;

[0188] Indicates the first The prediction variance of each quality indicator.

[0189] Formula (14): Physical consistency loss

[0190]

[0191] In the formula, Indicates the loss of physical consistency;

[0192] Indicates the number of constructed sample pairs;

[0193] , Represents any two samples;

[0194] , Representing samples respectively and samples A tiered risk index;

[0195] , Representing samples respectively and samples Predictive quality indicators;

[0196] This indicates the safety margin used to control the strength of the interval constraint.

[0197] Formula (15): Domain alignment loss

[0198]

[0199] In the formula, Indicates domain alignment loss;

[0200] Indicates the number of digital twin samples;

[0201] Indicates the number of actual diamond samples;

[0202] Indicates the first The fusion features of digital twin samples;

[0203] Indicates the first Fusion characteristics of a single real diamond sample;

[0204] This represents the eigenmap function that maps to the reproducing kernel Hilbert space;

[0205] Represents the regenerated nucleus Hilbert space.

[0206] Formula (16): Uncertainty calibration loss

[0207]

[0208] In the formula, Indicates the uncertainty calibration loss;

[0209] Indicates the number of predicted tasks;

[0210] Indicates the first The prediction variance of the output of each task;

[0211] Represents the actual measured value;

[0212] This represents the predicted mean;

[0213] This represents the true squared residual.

[0214] The technical effect achieved in this step is that the model can not only output accurate quality predictions, but also predictive uncertainties that can be used for risk assessment, and the prediction trend is subject to hierarchical risk physical prior constraints. The underlying mechanism is that multi-task prediction loss ensures numerical fitting, physical consistency loss limits the network prediction manifold from deviating from the physical laws of cutting, domain alignment loss reduces the distribution difference between digital twin samples and actual drilling samples, and uncertainty calibration loss makes the prediction variance correspond to the actual error magnitude. Therefore, even with a small sample size or changing operating conditions, the model is less likely to produce results that violate physical common sense, such as "increased feed rate but decreased axial force".

[0215] VI. Perform active learning supplementation and incremental model updates

[0216] The purpose of this step is to enable the model to adaptively adjust to tool wear, material batch variations, and operating condition expansion, rather than relying on one-time offline training. In online applications, for each candidate process parameter or actual drilling sample, sample novelty, maximum prediction uncertainty, and stratified risk index are calculated, forming an active learning evaluation metric.

[0217] Formula (17): Sample novelty

[0218]

[0219] In the formula, Indicates candidate samples The novelty of the sample;

[0220] This indicates that there is already a set of actual drill training samples;

[0221] This indicates the first [item] in the existing set of real-diamond training samples. One sample;

[0222] This represents the low-dimensional fusion representation of the candidate samples output by the cross-attention fusion layer;

[0223] This represents the low-dimensional fusion representation corresponding to historical actual drill samples.

[0224] This represents the Euclidean distance.

[0225] Formula (18): Active Learning Evaluation Indicators

[0226]

[0227] In the formula, Indicates candidate samples Active learning evaluation metrics;

[0228] This represents the maximum prediction standard deviation of the candidate sample across all prediction tasks.

[0229] Indicates the novelty of the sample;

[0230] This indicates a tiered risk index;

[0231] , , These represent the weighting coefficients for prediction uncertainty, sample novelty, and stratification risk, respectively.

[0232] When the active learning evaluation metric exceeds a preset threshold, the candidate parameter combination is added to the replenishment list and verified by actual drilling. New samples obtained from the replenishment are first entered into the incremental cache pool. When the number of samples in the cache pool reaches the preset micro-batch capacity, the system extracts historical samples covering different wear stages and parameter boundaries from the experience replay buffer and mixes them with the new samples for training. During incremental updates, the system preferentially freezes the underlying feature extraction layers of the static parameter branch and the dynamic temporal branch, only fine-tuning the cross-attention fusion layer and the multi-task output layer.

[0233] The technical benefits achieved in this step are as follows: the model can prioritize supplementing real-world drilling samples in areas of high uncertainty, high novelty, or high stratified risk, reducing a large number of low-value repetitive experiments, and suppressing catastrophic forgetting through experience replay. The underlying mechanism is that prediction uncertainty corresponds to insufficient model knowledge, sample novelty corresponds to insufficient feature space coverage, and the stratified risk index corresponds to potential machining damage risk; the combination of these three factors can locate key samples that have both information gain and engineering significance. Local fine-tuning retains the basic cutting features already learned by the network, adjusting only the fusion weights and output mappings related to new working conditions, thus balancing online adaptation speed and the stability of historical knowledge.

[0234] VII. Establish a constrained multi-objective Bayesian optimizer and output the optimal set of process parameters.

[0235] The purpose of this step is to transform the quality prediction results into recommended process parameters that can directly guide the machining process. After training, spindle speed, feed per revolution, ultrasonic frequency, and amplitude are used as candidate optimization variables. The optimization objective is to minimize at least two of the following: peak axial force, delamination factor, hole wall roughness, and burr height. Constraints such as a lower limit for material removal rate, an upper limit for delamination risk index, and an upper limit for prediction uncertainty are also set.

[0236] Formula (19): Constrained multi-objective Bayesian optimization criterion

[0237]

[0238] In the formula, This represents the candidate process parameters obtained through optimization;

[0239] Represents a vector of candidate process parameters to be evaluated;

[0240] Indicates the space of process parameters that are allowed to be searched;

[0241] Candidate parameters Contribution to the expected hypervolume increment of the target space;

[0242] Candidate parameters The joint probability that satisfies all constraints.

[0243] Formula (20): Joint probability feasibility

[0244]

[0245] In the formula, Candidate parameters The feasibility of joint probability;

[0246] , … They represent the 1st to the 1st. One constraint function;

[0247] Indicates the number of constraints;

[0248] Indicates probability operations;

[0249] Indicates the first One constraint is satisfied.

[0250] The optimizer dynamically modulates the acquisition function using joint probabilistic feasibility. When a candidate parameter is in the model blind zone, a high-risk stratified region, or its processing efficiency does not meet requirements, its feasibility probability decreases, and the acquisition function is penalized. Conversely, when a candidate parameter simultaneously possesses good quality prediction results, low uncertainty, and meets efficiency requirements, it is prioritized. The final output is a set of Pareto process parameters for process engineers to select based on quality-first, efficiency-first, or balanced strategies.

[0251] The technical effect achieved in this step is that the optimization result is no longer just theoretically optimal in terms of quality, but rather engineering-executable optimal under constraints of processing efficiency, physical risk, and model confidence. The underlying mechanism is that Bayesian optimization can efficiently explore the parameter space with a limited number of evaluations, while joint probabilistic feasibility transforms prediction uncertainty and hierarchical risk into a safety boundary during the optimization process, thereby preventing the algorithm from recommending parameters to regions where the model is unreliable or where there is a high physical risk.

[0252] In a typical implementation scenario, the test conditions were selected as follows: spindle speed of 6000 rpm, feed per revolution of 0.04 mm / rpm, ultrasonic frequency of 20 kHz, amplitude of 4 μm, and tool flank wear width of 0.1 mm. Dynamic signals were collected by a multi-source sensing system and input into a trained physical constraint temporal fusion network. The model output the predicted mean and corresponding predicted variance of peak axial force, exit delamination factor, hole wall roughness, and burr height. If the maximum predicted standard deviation is below the safety threshold and the delamination risk index is within the allowable range, the hole is processed according to the current process. If the predicted uncertainty enters the warning range, the control system can reduce the feed per revolution or add the condition to the replenishment list. If the predicted uncertainty remains high, manual inspection or offline non-destructive testing is triggered.

[0253] During the process planning phase, the lower limit for material removal rate is set at 300 cubic millimeters / minute, the upper limit for stratification risk index is set at 0.95, and the maximum prediction standard deviation is set to not exceed the warning threshold. A multi-objective Bayesian optimizer is constrained to output a set of Pareto optimal process parameters within the safety domain. For example, a spindle speed of 8500 rpm, a feed per revolution of 0.05 mm / rpm, an ultrasonic frequency of 25 kHz, and an amplitude of 3 micrometers can be selected as a balanced recommended point that considers both processing efficiency and quality. This recommended point ensures processing cycle time while reducing the risk of exit stratification and increased hole wall roughness.

[0254] The steps described above are not simply independent superpositions, but rather form a closed-loop collaborative relationship of "multi-source sensing—physical prior—virtual-real coupling—temporal fusion prediction—active learning—constraint optimization". The multi-source sensing system provides the actual processing status, the digital twin model provides physically interpretable boundaries, the static-dynamic hybrid sample library completes the fusion of physical samples and actual drilling samples, the temporal fusion network realizes online quality prediction, the uncertainty output provides a reliable basis for active learning and optimization constraints, and the Bayesian optimizer transforms the prediction results into executable process parameters.

[0255] The comprehensive technical effects of this collaborative mechanism include: First, improved prediction accuracy. This is because the model simultaneously utilizes static process parameters, tool status, dynamic process signals, and physical prior information, avoiding information gaps caused by a single data source. Second, improved stability across operating conditions. This is because the physical consistency loss and stratified risk index limit the model's output range, ensuring it still conforms to cutting physics even with small samples and varying tool wear conditions. Third, reduced experimental costs. This is because active learning only collects actual drilling samples in high information gain or high-risk areas, reducing repeated experiments in stable regions. Fourth, optimization results are more suitable for engineering implementation. This is because the optimization process simultaneously considers machining efficiency, quality indicators, physical risks, and prediction uncertainty, avoiding the generation of unexecutable or low-confidence recommended parameters.

[0256] From a fundamental perspective, the quality formation process of ultrasonic vibration-assisted drilling is essentially determined by kinematic contact state, thermo-mechanical coupling, tool wear, material delamination and fracture, and dynamic process response. This embodiment extracts kinematic and mechanical laws through a digital twin model, senses real-time states through sensor signals, learns difficult-to-analyze nonlinear residuals through neural networks, determines the model's confidence boundary through uncertainty estimation, and searches for optimal parameters within the confidence boundary using an optimizer. Therefore, this invention can achieve coordinated optimization of quality prediction and parameter recommendation under complex and variable working conditions.

[0257] The overall working principle of this invention can be summarized as follows: Before each ultrasonic vibration-assisted drilling, the system calls a rapid parameterized digital twin model based on the material to be processed, the tool state, and candidate process parameters to obtain axial force priors and layered risk priors; during the drilling process, the multi-source sensing system synchronously collects dynamic signals such as current, acoustic emission, vibration, and temperature, and inputs them together with static process parameters into a physical constraint temporal fusion network; the network explicitly models the modulation relationship between static working conditions and dynamic signals through a cross-attention mechanism, and outputs multiple drilling quality indicators and their prediction uncertainties.

[0258] When the prediction results are in a region with high confidence and low stratification risk, the system executes the machining according to the process parameters recommended by the constrained multi-objective optimizer. When the prediction uncertainty or stratification risk increases, the system incorporates the corresponding samples into the active learning and replenishment process, and corrects the network parameters through incremental updates. As the number of machining batches increases and tool wear evolves, the model continuously absorbs new actual drilling feedback, enabling the prediction model and the process optimizer to adapt to changes in on-site working conditions. Thus, this invention achieves a closed-loop operation from physical prior modeling, online quality prediction, risk identification, actual drilling replenishment, to parameter self-optimization.

Claims

1. A method for predicting the quality of ultrasonic vibration-assisted drilling of carbon fiber composite materials and optimizing parameters, characterized in that, Includes the following steps: Construct a multi-source sensing system for ultrasonic vibration-assisted drilling process to collect drilling process parameters, tool status information and dynamic signals of the drilling process. A rapid parametric digital twin model corresponding to the ultrasonic vibration-assisted drilling process is established. The rapid parametric digital twin model is used to characterize the effects of ultrasonic vibration, feed motion, thermal softening effect and tool wear on drilling load and exit delamination risk, and outputs the physical prior information of the drilling process. Based on parametric experimental design, the aforementioned physical prior information, and actual drilling measurement data, a static and dynamic hybrid sample library is constructed, and the deviation between the digital twin sample and the actual drilling sample is corrected. A physical constraint temporal fusion network is constructed, which includes a static parameter branch for processing static input features, a dynamic temporal branch for processing dynamic process signals, and a feature fusion module for fusing static features and dynamic temporal features. The physical constraint temporal fusion network is trained using the static-dynamic hybrid sample library to output the prediction results of borehole quality indicators and the prediction uncertainty corresponding to the prediction results. Based on the predicted uncertainty and the hierarchical risk information output by the fast parameterized digital twin model, active learning sampling and incremental model updates are performed. Based on the trained physical constraint temporal fusion network, a constrained multi-objective optimization model is established to optimize the process parameters of ultrasonic vibration assisted drilling and output the optimal set of process parameters that meet the constraints of processing efficiency and quality.

2. The method according to claim 1, characterized in that, The drilling process parameters collected by the multi-source sensing system include at least one of spindle speed, feed per revolution, ultrasonic frequency, and ultrasonic amplitude; the tool status information includes at least one of tool geometry parameters, tool coating type, and tool wear; and the dynamic process signals include at least one of spindle current signal, acoustic emission signal, vibration signal, and temperature signal.

3. The method according to claim 2, characterized in that, Preprocessing the dynamic process signal includes: Establish a unified time reference using machine tool control signals or spindle encoder signals; Timing alignment of dynamic process signals with different sampling rates; Smoothing interpolation processing is performed on low-frequency dynamic process signals; Extract envelope features, energy features, peak features, or statistical features from high-frequency dynamic process signals; The aligned dynamic process signal is segmented according to a sliding time window and then normalized to obtain multi-channel timing data for inputting the dynamic timing branch.

4. The method according to claim 1, characterized in that, The fast parameterized digital twin model includes: An ultrasonic kinematic model used to characterize the axial ultrasonic vibration displacement of a drill bit; A chip thickness model used to determine the instantaneous undeformed chip thickness based on the coupling relationship between feed motion and ultrasonic vibration; An intermittent contact model used to characterize the high-frequency contact and separation states between the tool and the workpiece; An axial force prior model is used to integrate the instantaneous undeformed chip thickness, intermittent contact state, thermal softening effect and tool wear effect to obtain the axial force prior time sequence. A stratification risk assessment model for determining export stratification risk based on axial force characteristics, material fracture properties, and laminate structural parameters during the export stage.

5. The method according to claim 4, characterized in that, The hierarchical risk assessment model determines the hierarchical risk index by considering the axial force impulse at the exit stage, the critical hierarchical thrust at the exit, the effect of ultrasonic load reduction, the effect of thermal softening, and the effect of tool wear. The hierarchical risk index is used as a physical consistency constraint in the training process of the physical constraint temporal fusion network, and as a risk constraint indicator in the active learning sampling and constrained multi-objective optimization process.

6. The method according to claim 1, characterized in that, The construction of the static-dynamic hybrid sample library includes: The initial process parameter combination is generated using a parameter space sampling method; The initial process parameter combination is simulated and calculated using the rapid parameterized digital twin model to obtain digital twin samples and their physical prior information; Collect a small number of actual drill samples and obtain the actual drill quality measurement labels corresponding to the actual drill samples; Based on the residual between the digital twin output and the actual drilling measurement results, a residual correction module is established to align the digital twin samples and the actual drilling samples in the virtual and real domains, thereby obtaining the static and dynamic hybrid sample library.

7. The method according to claim 1, characterized in that, In the physically constrained temporal fusion network: The static parameter branch is used to extract a global low-dimensional representation of process parameters, material layup information, and tool status information; The dynamic timing branch is used to extract local timing features and long-range dependence features of spindle current, acoustic emission, vibration and temperature; The feature fusion module employs a cross-attention mechanism to modulate the temporal features output by the dynamic temporal branch with the global low-dimensional representation output by the static parameter branch. The physical constraint temporal fusion network also includes a multi-task output layer, which is used to simultaneously output the predicted mean and prediction uncertainty of at least two of the following drilling quality indicators: peak axial force, layering factor, borehole wall roughness, and burr height.

8. The method according to claim 1, characterized in that, The physical constraint temporal fusion network is trained using a joint training objective, which includes: Multi-task prediction loss for fitting borehole quality measured labels; Physical consistency loss is used to constrain the relationship between prediction quality indicators and hierarchical risk information to maintain a physically monotonically consistent relationship. Domain alignment loss used to reduce the difference in feature distribution between digital twins and real diamond samples; Uncertainty calibration loss is used to constrain the prediction uncertainty to match the actual prediction error.

9. The method according to claim 1, characterized in that, The active learning imputation and incremental model update include: An active learning evaluation index is constructed based on the maximum prediction uncertainty, sample novelty, and stratified risk information; When the active learning evaluation index of the candidate sample meets the preset triggering conditions, the corresponding process parameter combination will be included in the actual drilling and sample replenishment list. The quality of the drill samples obtained from the supplementary drilling was measured and features were extracted, and then added to the incremental cache pool; When the number of samples in the incremental cache pool reaches the preset number, the new samples in the incremental cache pool are mixed with the historical samples in the experience replay buffer, and local network fine-tuning is performed. The local network fine-tuning includes freezing part of the basic feature extraction layer and updating the network parameters of the feature fusion module and the multi-task output layer.

10. The method according to claim 1, characterized in that, The constrained multi-objective optimization model takes minimizing at least two of the following indicators: peak axial force, layering factor, hole wall roughness, and burr height as the optimization objective, and uses at least one of the following constraints: lower limit of material removal rate, upper limit of layering risk, and upper limit of prediction uncertainty. The constrained multi-objective optimization model uses Bayesian optimization to search within the candidate process parameter space, and adjusts the acquisition function based on the probability that the candidate process parameters satisfy all constraints, outputting a Pareto process parameter set that satisfies the processing efficiency constraint and achieves optimal drilling quality.

Images