A parameter optimization method and system for machining of an inner drive guide rod end joint

Abstract

The application discloses a kind of parameter optimization method and system for machining end joint of inner drive guide rod, it is related to computer data processing and intelligent manufacturing technical field;Its technical points are as follows: the method includes sensing data acquisition and space-time dynamic feature extraction, dynamic routing selection and optimization of physical equation, construction and abstract granularity optimization of reversible causal abstract knowledge base, causal knowledge migration and decision fusion for new working condition, closed-loop evolution and knowledge distillation based on migration feedback;Through real-time acquisition of cutting signal by multiple sensors and combined with physical law to construct dynamic graph topology, space-time features of physical mechanism are extracted and fused using space-time graph convolution network, online perception and modeling of multi-physical field coupling dynamic characteristics are realized;Through physical equation dynamic routing, bidirectional gradient orthogonal projection protection and reversible causal abstract knowledge migration, self-adaptive suppression of chatter is realized, and causal knowledge in historical working condition is efficiently migrated to new working condition, to avoid repeated optimization.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and intelligent manufacturing technology, specifically to a parameter optimization method and system for machining the end joint of an internal drive guide rod. Background Technology

[0002] The internal drive guide rod end joint is a key transmission component used in high-speed precision stamping equipment. Its processing involves multiple steps such as turning, milling, and grinding, and the processing quality directly affects the motion accuracy and service life of the equipment. Due to the complex structure of the joint and the fact that the material is mostly high-strength alloy steel, the force-heat-vibration multi-physical field strong coupling during the cutting process can easily cause problems such as chatter, accelerated tool wear, and out-of-tolerance machining accuracy if the parameters are not selected properly.

[0003] Currently, there are two main technical approaches to optimize the machining parameters of the end joint of the internal drive guide rod. One approach is the empirical formula method based on metal cutting theory, which calculates and recommends cutting parameters using classical theoretical formulas such as the empirical model of cutting force and Taylor's tool life equation. The other approach is the statistical modeling method driven by sensor data, which collects multi-source sensor signals such as cutting force, temperature, and vibration, and uses machine learning or deep learning models to establish a mapping relationship between machining parameters and machining quality, thereby searching for the optimal parameter combination.

[0004] In the actual machining of the end joint of the internal drive guide rod, the workpiece material is usually tempered alloy structural steel, which has high hardness and poor thermal conductivity, making it easy for temperature to accumulate in the cutting zone. At the same time, the end joint is a typical thin-walled rotating body structure with weak radial rigidity, which easily induces chatter under cutting forces. Under the above conditions, the optimization of machining parameters faces two mutually restrictive technical difficulties: First, there is a strong coupling effect between cutting force, cutting temperature and system vibration. Increasing the cutting parameters will directly increase the cutting force. The cutting force will increase the cutting temperature through the Joule heating effect and plastic deformation. The increase in temperature will change the flow stress characteristics of the material and react on the cutting force. At the same time, the periodic fluctuation of the cutting force will excite the vibration of the process system, and the vibration will in turn modulate the amplitude of the cutting force. This strong coupling relationship of multiple physical fields makes it impossible for traditional independent optimization strategies based on a single physical quantity to obtain the global optimal solution. Second, the low stiffness of the thin-walled structure of the end joint makes the chatter threshold extremely sensitive to cutting parameters. When the combination of cutting parameters approaches the chatter boundary, weak precursor features in the vibration signal, such as subharmonic components in the spectrum and energy accumulation in specific frequency bands in the acoustic emission signal, are often submerged by strong background noise. Existing technologies cannot effectively identify these warning signals before chatter occurs, thus making it impossible to avoid chatter risks in advance during parameter optimization.

[0005] The combination of these two difficulties makes it difficult to effectively control the joint processing quality while ensuring processing efficiency through traditional methods for optimizing the processing parameters of the internal drive guide rod end joint. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for optimizing parameters in the machining of the end joint of an internal drive guide rod includes the following steps: Signals from the cutting process are acquired by multiple sensors and preprocessed. A theoretical correlation graph is generated based on physical laws, and a data-driven correlation matrix is ​​calculated. A dynamic graph topology matrix is ​​obtained through a gated fusion network, and spatiotemporal feature vectors are extracted. Based on the spatiotemporal feature vectors and the dynamic graph topology matrix, a physical equation branch library covering steady-state, chatter, and chip abrupt changes is constructed. A prototype equation structure is output through a hypernetwork routing generator. A balance between active exploration and utilization is introduced to select a physical equation branch. Parameter optimization is performed using bidirectional gradient coupling and orthogonal projection protection, resulting in a physical model confidence score. This confidence score is fed back to the gated fusion network for the next iteration. Based on the selected physical equation branch, microscopic physical variables... Causal inference is performed on macroscopic processing phenomena to construct microscopic and macroscopic causal graphs. Causal knowledge is extracted through reversible abstraction and refinement of triples, and the optimal abstract granularity parameters are obtained based on triple information bottleneck loss and Bayesian optimization. Based on the optimal abstract granularity parameters, reversible refinement expansion or dual-granularity collaborative transfer processing is performed on the target working condition. The prediction results of coarse-grained path and fine-grained path are weighted and fused through a fusion gating network to obtain the comprehensive prediction result under the target working condition. The prediction error is calculated based on the comprehensive prediction result and the actual processing result of the target working condition, and backpropagation is performed to update the trainable parameters of the forward link. The effective causal knowledge is distilled into the global knowledge base to complete the closed-loop update.

[0007] Furthermore, the process of acquiring and preprocessing signals from the cutting process using multiple sensors includes: acquiring a sensor signal set using multiple source sensors, wherein the sensor signal set includes at least: temperature signal, cutting force signal, acceleration signal, and acoustic emission signal; downsampling the sensor signal set to the same frequency; and applying different high-pass filter cutoff frequencies to different types of signals, wherein the cutoff frequency of the temperature signal is lower than that of other signals; the process of obtaining the dynamic graph topology matrix through a gated fusion network includes: calculating scalar fusion weights based on the physical model confidence score obtained from the previous iteration; and performing element-wise weighted summation of the theoretical correlation graph and the data-driven correlation matrix to obtain the dynamic graph topology matrix.

[0008] Furthermore, the physical equation branch library includes heat conduction equations and constitutive equations under steady-state cutting conditions, dynamic cutting force modulation equations under chattering conditions, and transient thermal shock equations and dynamic evolution equations of friction coefficients under abrupt changes in chip morphology. The hypernetic routing generator consists of fully connected layers, outputs multidimensional structure vectors and divides them into multiple sets of coefficients, which respectively adjust the adjustable parameters of each physical equation branch to generate the equation structure prototype.

[0009] Furthermore, the process of actively exploring and utilizing the balance to select a branch of a physical equation includes: by maintaining the equation structure value function, randomly selecting a branch of the physical equation with an exploration rate or by selecting the maximum value of the value function, and applying a Gaussian perturbation to the adjustable parameter coefficients of the selected branch's equation after each selection; the value function is updated in the form of an exponential moving average and has a minimum lower limit.

[0010] Furthermore, the process of performing parameter optimization by executing bidirectional gradient coupling and orthogonal projection protection includes: calculating the data gradient of the data fitting loss and the physical gradient of the physical constraint loss; if the dot product of the two is negative, the data gradient is projected onto a subspace orthogonal to the physical gradient, and only the projected data gradient is used to update the network parameters; the physical model confidence score is obtained by exponentially smoothing the root mean square error of the residuals of consecutive batches, taking the inverse, and then transforming it through an exponential function.

[0011] Furthermore, the causal inference process includes: applying historical maximum, historical minimum, and random offset perturbations to candidate causal variables, and performing forward inference under physical constraints to determine causal relationships; the reversible abstract refinement triplet includes a forward abstract encoder, a reverse refinement decoder, and a reversibility residual component, used to encode, reconstruct, and store difference information of the micro-causal graph subgraphs; the process of obtaining the optimal abstract granularity parameter includes: the forward abstract encoder encodes the subgraphs of the micro-causal graph to obtain the embedded representation of macro-concept nodes and outputs the information loss amount; the reverse refinement decoder reconstructs the micro-subgraphs based on the embedded representation and contextual features; the reversible residual component stores the difference information between the reconstructed subgraphs and the original subgraphs; Bayesian optimization aims to maximize the transfer effect function and iteratively searches to obtain the optimal abstract granularity parameter.

[0012] Furthermore, the dual-granularity collaborative transfer process includes: dividing the transfer strategy into three types based on the clustering results of the abstract granularity parameters of the historical task, including high-level abstraction transfer, medium-level abstraction transfer, and low-level abstraction transfer; the medium-level abstraction transfer simultaneously executes two transfer paths, including a coarse-grained path and a fine-grained path. The coarse-grained path predicts the probability distribution of macroscopic processing phenomena based on the macroscopic causal graph, while the fine-grained path predicts microscopic process results based on the reconstructed microscopic subgraph and converts them into macroscopic phenomenon probability values. The fine-grained path includes three output heads, which output the probability distribution of chatter, the probability distribution of chip morphology type, and the probability distribution of processing accuracy deviation level, respectively; the fusion gating network calculates the prediction uncertainty measure of the coarse-grained path and the fine-grained path, and assigns weights to perform a weighted summation of the prediction results of the two paths to obtain a comprehensive prediction result.

[0013] Furthermore, the process of distilling effective causal knowledge into the global knowledge base includes: storing the directed edge weights of the macro-causal graph that have been verified accurately, the new micro-causal relationships that have passed the significance test, and the verified abstract mapping relationships into the global knowledge base; when the average value of the transfer effect function of multiple consecutive transfer tasks is greater than the preset average value threshold and the fluctuation of the optimal abstract granularity parameter is less than the preset standard deviation threshold, knowledge transfer is performed directly without the need for a small number of samples to pre-optimize.

[0014] A parameter optimization system for machining the end joint of an internally driven guide rod includes: a sensor signal preprocessing and dynamic topology construction module, which resamples, aligns, and differentially filters the cutting signals acquired by multiple sensors, generates a theoretical correlation graph and calculates a data-driven correlation matrix, fuses the two based on the confidence score from the previous iteration to obtain a dynamic graph topology matrix, and extracts spatiotemporal feature vectors through spatiotemporal graph convolution; a physical constraint optimization and confidence assessment module, which constructs an equation branch library based on the spatiotemporal feature vectors and the dynamic graph topology matrix, generates an equation structure prototype through a hypernetwork and selects an equation branch using an exploration-based balancing mechanism, performs orthogonal projection protection on conflict gradients during backpropagation, calculates the confidence score after optimization, and feeds it back to the sensor signal preprocessing and dynamic topology construction module; causal knowledge The abstraction and granularity optimization module constructs a micro-causal graph by performing counterfactual intervention tests on micro-physical variables based on the selected equation branch, and also constructs a macro-causal graph. It performs lossy abstraction and refinement through forward abstraction, reverse refinement, and residual memory, and obtains the optimal abstraction granularity parameters based on Bayesian optimization. The knowledge transfer and decision fusion module selects a transfer strategy based on the optimal abstraction granularity parameters, performs lossy refinement expansion or dual-granularity collaborative transfer on the target working condition, and weights and fuses the results of coarse-grained and fine-grained paths through a gating network to obtain a comprehensive prediction result. The closed-loop evolution and knowledge distillation module calculates the prediction error based on the actual processing results and the comprehensive prediction result, backpropagates it, updates trainable parameters, distills effective causal knowledge into the global knowledge base, and manages the elimination queues of different types of data.

[0015] Furthermore, in the sensor signal preprocessing and dynamic topology construction module, resampling alignment unifies all sensor signals to the same frequency; differential filtering uses high-pass filtering for cutting force, acceleration, and acoustic emission signals, and low-pass filtering for temperature signals; in the physical constraint optimization and confidence assessment module, the equation branch library includes physical equations under steady-state, chatter, and chip abrupt change states; the exploration utilizes a balance mechanism to maintain the value function and selects branches with a decaying exploration rate; orthogonal projection protection uses only the projected data gradient to update parameters in the event of gradient conflict; in the causal knowledge abstraction and granularity optimization module, counterfactual intervention testing perturbs each candidate causal variable with historical maximum, minimum, and random offset perturbations, and determines causal relationships based on the amount of change; lossy abstraction refinement encodes micro-subgraphs into macro-embedded structures using a forward encoder. The system outputs information loss and reconstructs micro-subgraphs based on macro-embedding and contextual features using an inverse decoder. Reconstructed difference information is stored using a residual memory component. In the knowledge transfer and decision fusion module, dual-granularity collaborative transfer includes coarse-grained and fine-grained paths. The coarse-grained path outputs macro-phenomenal probabilities based on the macro-causal graph, while the fine-grained path converts micro-quantity predictions into macro-phenomenal probabilities for three output heads using an aggregation function. A gating network assigns fusion weights based on prediction uncertainty. In the closed-loop evolution and knowledge distillation module, the global knowledge base maintains independent least recently used elimination queues for causal edges, equation parameters, and abstract mappings. When the average effect function and optimal abstract granularity parameter fluctuations of multiple consecutive transfer tasks meet the convergence threshold and the prediction error is lower than the accuracy threshold, the system switches to direct knowledge transfer mode.

[0016] This invention provides a method and system for optimizing parameters in the machining of the end joint of an internal drive guide rod, which has the following beneficial effects: 1. This invention utilizes the physical laws of metal cutting to generate a theoretical correlation graph as a priori constraint, while simultaneously calculating a data-driven correlation matrix from real-time sensor data. The two are dynamically fused into a graph topology using scalar fusion weights guided by confidence scores, enabling the deep learning model to obtain a reasonable physical inductive bias in the early stages of training. Furthermore, a bidirectional gradient coupling and orthogonal projection protection mechanism based on a physical information neural network is introduced to ensure that the data fitting process does not destroy the integrity of physical laws in the parameter update direction, thereby achieving a dynamic balance between physical consistency and data adaptability under strong force-heat-vibration coupling conditions. In addition, counterfactual intervention testing and reversible abstract refinement of triples extract causal knowledge transferable across working conditions, enabling the optimized model to maintain stable prediction accuracy under different processing conditions, fundamentally avoiding the shortcomings of insufficient generalization ability in purely data-driven methods.

[0017] 2. This invention employs a multi-source sensor signal resampling alignment and differential filtering preprocessing strategy. For cutting force, acceleration, and acoustic emission signals, high-pass filtering is used to extract dynamic components, while temperature signals retain their slow-varying characteristics, ensuring the physical rationality of subsequent deep learning feature extraction. In the spatiotemporal feature extraction stage, a dynamic graph topology matrix is ​​used as the graph structure, and multi-sensor time-series signals are used as node features. A spatiotemporal graph convolutional network extracts deep spatiotemporal feature vectors that fuse spatial topological relationships and temporal evolution laws, enabling the deep learning model to effectively extract flutter precursor features from sensor signals submerged in strong background noise. In the causal knowledge transfer stage, a micro-causal graph is constructed through counterfactual intervention testing, and causal knowledge transfer across operating conditions is achieved through reversible abstraction and refinement of triples. This allows flutter warning knowledge accumulated in the source operating condition to be effectively transferred to the target operating condition, thereby proactively avoiding flutter risks during parameter optimization.

[0018] 3. This invention establishes a complete closed-loop update mechanism from backpropagation of prediction error to all trainable parameters in the forward link, and continuously distills verified causal knowledge into the global knowledge base, enabling the optimization model to continuously evolve across working conditions. When the model converges and switches to zero-sample transfer mode, knowledge transfer and parameter optimization can be directly performed without collecting sample data of the target working condition, significantly reducing the process debugging cost and time cycle under the new working condition. Overall, it organically integrates the determinism of physical mechanisms with the data-driven power of deep learning, and constructs an intelligent optimization system for cutting parameters that combines physical interpretability and data adaptability. It can be widely applied to CNC cutting parameter optimization scenarios for various precision rotating parts. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention.

[0020] Figure 2 This is a diagram illustrating the dynamic map construction and spatiotemporal feature extraction of the present invention.

[0021] Figure 3 This is a diagram illustrating the migration route branching and closed-loop feedback of the present invention.

[0022] Figure 4 This is a system framework diagram of Embodiment 2 of the present invention. Detailed Implementation

[0023] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1: Please see Figures 1 to 3 This embodiment provides a parameter optimization method for machining the end joint of the internal drive guide rod, including steps 1 to 5.

[0025] Step 1: Multi-source sensor data acquisition and spatiotemporal dynamic feature extraction guided by physical means: Multi-source sensor signals are acquired in real time during the cutting process of the end joint of the internal drive guide rod. After preprocessing, the sensor correlation graph is dynamically constructed by combining prior knowledge generated by the physical laws of metal cutting with the statistical characteristics of real-time data. Then, a time-series graph convolutional network is used to extract the spatiotemporal feature vector of the deep fusion physical mechanism.

[0026] Step 1.1: Real-time acquisition and preprocessing of multi-source sensor signals.

[0027] The following sensors are installed on the spindle, tool post, and workpiece fixture of the CNC machine tool: a triaxial cutting force sensor to acquire the components of the cutting force in three orthogonal directions, with a sampling frequency of 5000 Hz, sufficient to capture the main dynamic components of the cutting force; a thermocouple sensor to acquire the temperature of the cutting zone, with a sampling frequency of 1000 Hz, satisfying the slow variation characteristics of temperature changes; an accelerometer to acquire spindle vibration signals, with a sampling frequency of 10000 Hz, covering the common chatter frequency range; and an acoustic emission sensor to acquire high-frequency stress wave signals, with a sampling frequency of 20000 Hz, used to detect microscopic events such as tool wear and chip breakage. Since the sampling frequencies of each sensor are different, resampling is performed first: all sensor signals are uniformly resampled to 20000 Hz to ensure strict temporal alignment; all sensor signals are acquired synchronously, with each data acquisition batch lasting 5 seconds, covering at least one complete tool cutting stroke.

[0028] The acquired raw signals are preprocessed: differentiated filtering strategies are adopted for signals from different sensors; for the three-dimensional cutting force signal, acceleration signal, and acoustic emission signal, a high-pass filter with a cutoff frequency of 50 Hz is used to remove low-frequency drift noise. The selection of this cutoff frequency is based on the fact that the feed frequency in the cutting force signal is usually below 20 Hz, while the main dynamic component of the cutting force is above 50 Hz. The 50 Hz high-pass filter can effectively suppress low-frequency thermal drift and baseline drift caused by machine tool structural deformation while retaining the main dynamic component; for the temperature signal, since its energy is concentrated in the extremely low frequency, usually below 1 Hz, no high-pass filtering is performed. Only moving average smoothing or low-pass filtering is used to remove high-frequency noise. The cutoff frequency is set to 5 Hz to retain the slow-changing characteristics of the cutting temperature.

[0029] At each sampling point, a sliding window with a length of 200 sampling points is taken with the current point as the center. The root mean square value of the signal within the window is calculated as the effective value at that moment, thus preserving the complete time series. The preprocessed result is a four-dimensional feature vector. The four dimensions correspond to the three-dimensional cutting force composite amplitude, cutting temperature, vibration amplitude, and acoustic emission energy root mean square value, respectively. The three-dimensional cutting force composite amplitude is calculated using the Euclidean norm.

[0030] Step 1.2: Generate a theoretical correlation diagram based on the physical laws of cutting.

[0031] Based on the force-thermal-vibration coupling relationship in metal cutting principles, a theoretical correlation graph is constructed. This graph contains four nodes, corresponding to the cutting force node, temperature node, vibration node, and acoustic emission node, respectively. The weights of the directed edges between nodes are pre-set according to the physical propagation laws: the weight from cutting force to temperature is 0.8, indicating that the cutting force has a strong positive influence on the cutting temperature, based on the empirical ratio of Joule heating effect and plastic deformation heat; the weight from cutting force to vibration is 0.6, reflecting the degree of effect of cutting force fluctuation on vibration excitation; the weight from temperature to vibration is 0.4, indicating the indirect influence of thermal deformation on vibration modes; the weight from acoustic emission to cutting force is 0.5, characterizing the feedback of microscopic fracture events to macroscopic cutting forces; the weight between the remaining nodes is 0.2 as baseline coupling; the weight of the self-loop of a node is uniformly set to 0, indicating that the direct influence of the variable on itself is not reflected through the graph structure; this theoretical correlation graph is stored as a physical prior matrix, with a matrix size of 4×4, and each element value is the weight of the corresponding edge.

[0032] Step 1.3: Calculate the data-driven correlation matrix of the real-time signal.

[0033] Based on the four-dimensional feature vector, a time window is defined as every 200 consecutive sampling points, with a window sliding step of 50 sampling points. This setting ensures that each window covers a sufficient amount of dynamic changes while keeping the computational load moderate. Within each time window, the Pearson correlation coefficient between the four feature channels is calculated, resulting in a 4×4 correlation coefficient matrix as the data-driven matrix. The data-driven matrix has the same matrix dimension as the theoretical correlation graph.

[0034] Step 1.4: Dynamically fuse theoretical correlation graphs and data-driven matrices through gated fusion networks.

[0035] A physical model confidence score is introduced, with a value ranging from 0 to 1. This score is calculated by step 2.4 in the previous complete iteration and stored in the history. For the first iteration, since there is no history, the initial confidence score is determined by the following rules: if it is the first run and there is no historical data, the initial score is set to 0.5; if there is historical data of the source condition but the current target condition is completely new, the initial score is set to 0.8.

[0036] The scalar fusion weight α is calculated based on the confidence score, according to the following rules: When the confidence score is greater than 0.7, α equals 1, representing the complete trust theory correlation graph; When the confidence score is <0.3, α equals 0, indicating complete trust in the data-driven matrix; When 0.3 ≤ confidence score ≤ 0.7, α equals confidence score minus 0.3 divided by 0.4, that is, α increases linearly with confidence score.

[0037] After obtaining the scalar α, the dynamic graph topology matrix is ​​calculated as follows: the dynamic graph topology matrix is ​​equal to α multiplied by the theoretical association graph matrix, plus 1 minus α multiplied by the data-driven matrix. All operations are element-wise multiplications.

[0038] This approach ensures that the dynamic graph topology matrix is ​​a simple convex combination of the theoretical matrix and the data matrix, with each element automatically falling between 0 and 1, and without the problem of quadratic weighting; the physical meaning of α is clear: when the physical model has high credibility, it relies more on theoretical priors; when the credibility is low, it relies more on real-time data.

[0039] Step 1.5: Extract deep spatiotemporal feature tensors using a spatiotemporal graph convolutional network.

[0040] Using a dynamic graph topology matrix as the graph structure and the preprocessed four-dimensional feature vectors from step 1.1 as node features, a spatiotemporal graph convolutional network is constructed. This network contains two spatiotemporal convolutional blocks, each of which contains a graph convolutional layer and a temporal convolutional layer. The graph convolutional layer uses Chebyshev polynomial approximation with an order of 3 to capture the interaction of sensors within the local neighborhood. The temporal convolutional layer has a kernel size of 5, a stride of 2, and 64 output channels to extract the temporal evolution pattern of the signal. After the two spatiotemporal convolutional blocks, a global average pooling layer and a fully connected layer are connected. The fully connected layer has an output dimension of 128.

[0041] The input to the spatiotemporal graph convolutional network is a three-dimensional tensor with the shape of the number of time windows × the number of nodes × the feature dimension. The number of time windows is calculated as: the total number of sampling points minus the window length, divided by the sliding step size, and then added by 1. In this method, the single batch acquisition time is 5 seconds, the sampling frequency of each sensor is known, and the corresponding total number of sampling points can be calculated. All sensor signals are acquired synchronously, and the same formula is used. The output is a spatiotemporal feature vector with a dimension of 128. This feature vector integrates the spatial topological relationship and temporal evolution law of the sensor signals, and the physical mechanism prior has been incorporated through dynamic graph topology.

[0042] Step 2: Dynamic routing and safety optimization of physical equations: Establish a branch library of physical equations covering various cutting physical states, generate parameterized equation structure prototypes through differentiable supernetworks, introduce multi-armed gambling machine-style exploration to utilize balance mechanisms and gradient orthogonal projection protection, and dynamically optimize physical model parameters.

[0043] Step 2.1: Construct a branch library of physical equations.

[0044] A library of equations containing five branches of physical equations is pre-built, based on typical physical states that may occur during the machining of the end joint of the drive rod within the coverage area, such as steady state, chatter, and chip abrupt changes.

[0045] Branch 1 is the heat conduction equation under steady-state cutting conditions, in the form of: partial derivative of cutting temperature with respect to time = thermal diffusivity × divergence of temperature gradient; where thermal diffusivity = thermal conductivity / (density × specific heat capacity), and the thermal diffusivity is determined by the workpiece material properties; the divergence of temperature gradient is calculated from the temperature field distribution using the finite difference method.

[0046] Branch two is the Johnson-Cook constitutive equation under steady-state cutting conditions, in the form: Flow stress = (Reference yield stress + Hardening coefficient × Equivalent plastic strain) 应变硬化指数 ) × (1 + strain rate sensitivity coefficient × logarithmic dimensionless strain rate) × (1 - dimensionless temperature) 热软化指数 ); where, dimensionless strain rate = current strain rate / reference strain rate, dimensionless temperature = (current temperature - room temperature) / (melting temperature - room temperature); the reference yield stress ranges from 200 to 1500 MPa, the hardening coefficient ranges from 100 to 1000 MPa, the strain hardening index ranges from 0.1 to 0.7, the strain rate sensitivity coefficient ranges from 0.01 to 0.1, and the thermal softening index ranges from 0.5 to 1.5. In actual use, matching parameters are selected within the corresponding range according to the workpiece metal grade, tool material, and machining conditions; the lower limit of the range is selected for soft metals such as aluminum alloys, and the upper limit of the range is selected for high-strength hard metals such as alloy steel.

[0047] Branch 3 presents the dynamic cutting force modulation equation under chatter conditions, in the form: Dynamic cutting force = Static cutting force + Cutting force coefficient × Cutting width × (Current vibration displacement - Previous rotation vibration displacement); where the current vibration displacement and the previous rotation vibration displacement are obtained by integration from the accelerometer, and the cutting force coefficient is determined by the workpiece material and tool geometry, with a value ranging from 500 to 3000 N / mm. 2 .

[0048] Branch four presents the transient thermal shock equation for abrupt changes in chip morphology, in the form: Heat flux density = -thermal conductivity × temperature gradient + transient shock heat source term; where the negative sign indicates heat transfer from the high-temperature region to the low-temperature region; the transient shock heat source term is generated by the heat released during chip fracture or abrupt morphology change, and is expressed as the initial heat source intensity multiplied by an exponential decay function, with the time when the exponent is negative divided by the decay time constant, where the initial heat source intensity ranges from 1000 to 5000 W / m. 2 The decay time constant ranges from 0.01 to 0.1 s.

[0049] Branch 5 presents the dynamic evolution equation of the friction coefficient under abrupt changes in chip morphology, in the form: Dynamic friction coefficient = Basic friction coefficient + Friction enhancement coefficient × (Vibration velocity / Reference velocity). 2 Among them, the basic friction coefficient is the friction coefficient between the tool and the workpiece material when sliding at low speed and in a stable manner, and the value ranges from 0.1 to 0.5; the reference speed is taken as 1 m / s for dimensionless processing; the friction enhancement coefficient is a dimensionless constant, and the value ranges from 0.01 to 0.2; the vibration speed is obtained by real-time measurement by the vibration sensor.

[0050] Each branch of the equation corresponds to a unique index number.

[0051] Step 2.2: Output the prototype equation structure through the hypernetwork probabilistic routing generator.

[0052] A hypernetwork routing generator is constructed, which consists of three fully connected layers. The first layer is the input layer with an input dimension of 128, consistent with the spatiotemporal feature vector output in step 1.5. The second layer is the hidden layer with a dimension of 256, using a linear rectified function as the activation function. The third layer is the output layer with an output dimension of 50. The output 50-dimensional structure vector is divided into five groups, each containing 10 coefficients, corresponding to five branches of the physical equations.

[0053] The decoder receives 10 coefficients per group, which are used to adjust the adjustable parameters of the corresponding equations in step 2.1, such as scaling or shifting the reference yield stress and hardening coefficient in the Johnson-Cook equation, thereby generating a parameterized equation structure prototype. The decoder represents the adjustable parameter terms in each equation branch as a linear combination of basis functions, while the overall structure of the equation remains unchanged. The basis functions are a predefined set of functions, specifically: basis function one is a constant function with a value of constant one; basis function two is a linear function with a value of the independent variable itself; basis function three is a square function with a value of the square of the independent variable; basis function four is an exponential function with a value of the natural constant e raised to the power of the independent variable; and basis function five is a sine function with a value of the sine of the independent variable. The expression for each equation branch is equal to the first 5 coefficients multiplied by the corresponding basis function, and the last 5 coefficients are used for scaling and shifting the independent variable within the basis function. The five equation branch expressions output by the decoder constitute the set of equation structure prototypes.

[0054] Step 2.3: Introduce active exploration and utilize the equilibrium mechanism for equation structure selection.

[0055] Maintain an equation structure value function, which is an array of size 5, with all initial values ​​set to 0.5. At each decision, generate a random number between 0 and 1. If the random number is less than the exploration rate ε, randomly select one of the five equation branches as the activation branch, and apply a Gaussian perturbation with a mean of 0 and a standard deviation of 0.1 to the basis function coefficients corresponding to the activation branch to generate a new equation structure variant, based on ensuring a moderate exploration step size and avoiding drastic structural oscillations. If the random number is greater than or equal to ε, select the equation branch with the largest value in the current equation structure value function as the activation branch. The initial value of ε is set to 0.3 to ensure sufficient exploration in the early stages. After each complete iteration of steps 1 to 5, ε is multiplied by a decay factor, which is set to 0.99 with a lower bound of 0.05, based on the principle that classic reinforcement learning exploration uses a balanced strategy to ensure sufficient exploration in the early stages and converge to the optimal strategy in the later stages.

[0056] The selected equation branch serves as the physical constraint for the current operating condition and is used for subsequent training of the physical information neural network. After training, the equation structure value function is updated based on the prediction accuracy and physical residual reduction magnitude of the equation branch on the validation set: the new value function value equals the original value function value multiplied by 0.8 plus the normalized performance score multiplied by 0.2, where the performance score equals the prediction accuracy multiplied by 0.6 plus the physical residual reduction magnitude multiplied by 0.4; the prediction accuracy equals 1 minus the quotient of the validation set root mean square error divided by the initial root mean square error of the training set, with the result truncated to the interval between 0 and 1; the physical residual reduction magnitude equals the initial... The initial physical residual is subtracted from the current physical residual and then divided by the initial physical residual, with the result truncated to the interval between 0 and 1. The calculated performance score is then normalized using the minimum-maximum normalization method, i.e., the performance score is subtracted from the historical minimum performance score and then divided by the historical maximum performance score minus the historical minimum performance score, resulting in a normalized value between 0 and 1. If the historical record is empty, the original performance score is used directly as the normalized value. To prevent the value function from decaying to zero after long-term operation, the minimum value of all values ​​of the value function is set to 0.2 after each update. If the value function of a certain branch is lower than 0.2, it is forcibly reset to 0.2.

[0057] This update mechanism balances historical experience with the latest performance, and the selected equation branch and its current parameters are saved in a cache for use in the counterfactual intervention test in step 3.1.1.

[0058] Step 2.4: Perform bidirectional gradient coupling and orthogonal projection protection.

[0059] The equation branch selected in step 2.3 is used as a physical constraint term, which together with the data fitting loss constitutes the total loss function. The total loss function = data fitting loss + physical constraint loss × balance coefficient. The balance coefficient is initially 0.1 and increases by 0.04 every 100 iterations during training until it reaches 0.5 in the 1000th iteration. The basis for this is that the focus in the early stage of training is on data fitting to achieve fast convergence, while the physical constraints are gradually strengthened in the later stage to ensure physical rationality. If the iteration period exceeds 1000, the balance coefficient remains unchanged at 0.5.

[0060] During backpropagation, the gradient vector of the data fitting loss with respect to the network parameters is calculated and denoted as the data gradient; at the same time, the gradient vector of the physical constraint loss with respect to the network parameters is calculated and denoted as the physical gradient; the dot product of the data gradient and the physical gradient is calculated. If the dot product is negative, it indicates that the two gradient directions conflict; at this time, the data gradient is projected onto a subspace orthogonal to the physical gradient: the projected data gradient is equal to the data gradient minus the projection component of the data gradient in the direction of the physical gradient.

[0061] The magnitude of the physical gradient is equal to the square root of the sum of the squares of all elements of the physical gradient vector. The projection component is calculated by dividing the dot product of the data gradient and the physical gradient by the square of the magnitude of the physical gradient, and then multiplying by the physical gradient. If the dot product is positive, no projection operation is performed. Then, only the projected data gradient is used to update the network parameters, and the original physical gradient is no longer added.

[0062] The parameter update step size and learning rate are set to 0.001, and the momentum parameter is set to 0.9. These values ​​are the default parameters for the mainstream Adam optimizer, ensuring convergence stability. This operation ensures that the direction of network parameter updates does not disrupt the physical laws of the current activation, while preserving the data-driven learning capability.

[0063] After the physical information neural network is trained, the root mean square error of the prediction residuals of the currently selected equation branch is calculated on an independent validation set. An exponential smoothing filter is used to filter the root mean square errors of the residuals of five consecutive batches, with a smoothing factor set to 0.9. The filtered residual values ​​are converted into confidence scores, which are equal to e raised to the power of the negative residual, where e is a natural constant and the residual is the root mean square error after filtering. This conversion ensures that the score is 1 when the residual is zero and 0 when the residual approaches infinity. This score is combined with the feature vector of the current working condition to form a key-value pair and stored in the historical confidence score database.

[0064] Step 3: Construction and Optimization of the Reversible Causal Abstract Knowledge Base: Using the physical equations and counterfactual intervention tests obtained in Step 2, the causal knowledge base is constructed and updated, and the abstract granularity parameters are optimized. Due to information loss and the introduction of external features, the reversibility in this step is actually a lossy approximation.

[0065] Step 3.1: Construct micro-causal graphs and macro-causal graphs.

[0066] Step 3.1.1: Construct a micro-causal graph.

[0067] For six microscopic physical variables—cutting force amplitude, temperature gradient, peak frequency of vibration spectrum, acoustic emission energy rate, estimated tool wear, and estimated work-hardened layer depth—the following counterfactual intervention test was performed for each pair of candidate causal and candidate outcome variables: First, a baseline prediction state is established using the current physical equation branch selected in step 2.3.

[0068] The tool wear estimate is estimated online using a model that compares the energy accumulation of the acoustic emission signal with the cutting force ratio; the work-hardened layer depth estimate is calculated based on post-processing calculations using cutting temperature history and the Johnson-Cook constitutive model.

[0069] Secondly, for each candidate causal variable, at least 30 independent perturbation tests are conducted. The specific perturbation is allocated as follows: 10 tests for historical maximum value, 10 tests for historical minimum value, and 10 tests for random offset. The historical maximum value, historical minimum value, and historical standard deviation are all taken from all valid processed data collected in step one under the current operating condition. If it is the first iteration, the statistical value in the source operating condition knowledge base is used. The magnitude of the random offset is 1.5 times the historical standard deviation of the variable, and the positive and negative directions are randomly selected.

[0070] Next, forward extrapolation under physical constraints is performed: the combination of variables after intervention is input into the current physical equation branch selected in step 2 to recalculate the predicted values ​​of the result variables; this step ensures that the extrapolation results after variable changes strictly follow the physical laws of the current working conditions, rather than being a pure data statistical extrapolation.

[0071] Then, the observation results and causal determination are performed: the change in the predicted value of the outcome variable before and after the perturbation is calculated; if the absolute value of the change exceeds 0.5 times the historical standard deviation of the outcome variable, the candidate causal variable is determined to have a significant causal relationship with the outcome variable; the 0.5 times threshold is derived from two hundred sets of preliminary experimental data, which can ensure a true positive rate of over 95% while controlling the false positive rate below 5%; at this point, a directed edge from the causal variable to the outcome variable is added to the micro-causal graph; the number of perturbations is increased to 30 times, and when performing a significance test on the new edge in step 5.3, a t-test can be used directly, as the sample size is sufficient.

[0072] Finally, calculate the causal edge weights: repeat the above process for all variable pairs; ultimately, the weight of each edge is set to the normalized causal effect strength, which is the average change of the outcome variable in all perturbation tests, divided by the largest average change among all confirmed edges, so that the weight values ​​of all edges strictly fall between 0 and 1; if the average change of all edges is 0, then the weight of all edges is set to 0.

[0073] Step 3.1.2: Construct a macro-causal graph.

[0074] The node set of the macro-causal graph contains three macro-processing phenomenon nodes: chatter intensity index, chip morphology category, and processing accuracy deviation level. The directed edges of the macro-causal graph are initialized based on the expert knowledge base: for example, the weight of the edge pointing from the chatter intensity index to the processing accuracy deviation level is initially set to 0.7, and the weight of the edge pointing from the chip morphology category to the chatter intensity index is initially set to 0.5, based on the prior estimate of the causal intensity of macro-phenomena in the expert knowledge base. The initial structure of this macro-causal graph will be automatically adjusted and optimized in subsequent steps through a co-evolutionary mechanism. Specifically, step 5.3 supplements the update rules of the macro-causal graph: when the error of the macro-causal graph prediction on the validation set decreases by more than a threshold, directed edges are added or reversed, and the conditional probability table is updated using the Bayesian structure learning algorithm.

[0075] Step 3.2: Construct reversible abstract-refined triples.

[0076] Construct three learnable components to form a reversible abstract-refined triplet.

[0077] The first component is a forward abstract encoder, implemented using a graph neural network. The input to the graph neural network is a subgraph from a micro-causal graph, obtained through the following sampling strategies: a random walk strategy is used, starting from each node and sampling 3 to 5 neighboring nodes according to edge weights to generate a candidate subgraph set; alternatively, a graph attention network can be used to learn subgraph importance and select the subgraph with the highest mutual information. The graph neural network uses three concatenated graph convolutional layers, each outputting a 64-dimensional vector, followed by a global pooling layer to output a 128-dimensional vector. This vector is mapped through a fully connected layer to the embedding representation of macroscopic concept nodes, with a dimension of 32. The encoder outputs an information loss, which is obtained by calculating the mutual information difference between the micro-subgraph node features and the macro-node embedding representations. The mutual information is estimated by a discriminant network. This discriminant network is a two-layer fully connected network. The first layer has an input dimension of 32 plus 6 equals 38 and an output dimension of 16, and is activated using a linear rectified function. The second layer has an output dimension of 1, and the matching probability is obtained using a sigmoid function. During training, positive samples are pairs of true macro-node embeddings and corresponding micro-subgraph node features, and negative samples are shuffled pairs, optimized using a binary cross-entropy loss function. The value of the information loss is between 0 and 1.

[0078] The second component is the inverse refinement decoder, which is implemented by a conditional variational autoencoder. Its inputs are the embedded representations of macroscopic concept nodes (32-dimensional) and the contextual feature vectors of the target working condition (128-dimensional). The output is the probability distribution of the micro-subgraph. Specifically, the inverse refinement decoder first maps the inputs through a fully connected layer to a Gaussian distribution with a mean of 128-dimensional vectors and a log-variance of 128-dimensional vectors. Then, it samples a 128-dimensional latent vector from this Gaussian distribution. Finally, the graph decoder reconstructs the micro-subgraph from the latent vector. The graph decoder first generates a fixed set of node identifiers, corresponding to six fixed variable types in the micro-causal graph. Then, the latent vectors are mapped to the initial embeddings of each node through a fully connected layer. The embedding dimension of each node is set to 64, forming a 6x64 node feature matrix, which is then refined by two graph convolutional layers. The edge prediction layer is implemented by a fully connected layer, with the input being the concatenated vector of two node embeddings and the output being the probability of the existence of an edge between 0 and 1. The graph decoder outputs an adjacency matrix representing the structure of the micro subgraph. The training objective of this decoder is to maximize the structural similarity between the reconstructed micro subgraph and the original micro subgraph, which is measured using cross-entropy loss. Due to the existence of information loss and the introduction of working condition features, the reconstructed subgraph cannot be equal to the original subgraph. Therefore, strict invertibility is not required; only approximate recovery is required.

[0079] The third component is the reversible residual component, which is implemented by a residual memory network. The reversible residual component receives the information loss and the difference map between the reconstructed micro-subgraph and the original micro-subgraph. The difference map is obtained by comparing each edge, and the difference value of each edge is 0 or 1, indicating that they are the same or different. The difference map is first represented as a flattened vector of the adjacency matrix, and then mapped to a 128-dimensional embedding vector through a fully connected layer. The reversible residual component stores this difference information in a key-value pair memory matrix with a size of 1000×128, where 128 is the dimension of the difference map embedding vector. The update rule of the reversible residual component is as follows: when the cosine similarity between the new difference map embedding and the existing entries in the memory matrix is ​​less than 0.6, the new difference map embedding is added to the memory matrix as a new entry; otherwise, the most similar entry is updated with the new difference map using an exponential moving average, with an update coefficient of 0.9, based on the principle that the memory network can absorb new features while maintaining the stability of historical knowledge.

[0080] Step 3.3: Define the differentiable abstract granularity parameter and the triple information bottleneck loss.

[0081] The abstract granularity parameter λ is defined as a continuously differentiable scalar variable, ranging from 0 to 1, with an initial value of 0.5. A larger λ value indicates a higher level of abstraction, where more details in the micro-causal graph are compressed into macro-concept nodes; a smaller λ value indicates the retention of more micro-details. The mechanism of λ is as follows: In the forward abstract encoder, λ acts as a threshold to control the granularity of subgraph selection. When λ > 0.5, only nodes directly and strongly correlated with macro-phenomena are selected, such as cutting force, temperature, and vibration, while detailed nodes like tool wear are discarded. When λ < 0.5, more nodes are retained. When λ = 0.5, a balancing strategy is adopted, retaining all nodes but normalizing the sampling probability of each node by multiplying its importance score by 0.5, thus avoiding bias towards either macro-phenomenon nodes or micro-detail nodes. The importance score of a node multiplied by λ is used as the sampling probability. The gradient information of λ can be used as an aid to Bayesian optimization.

[0082] The total loss function is the sum of the three losses.

[0083] The first term is the prediction fidelity loss: On the validation set, the processing results are predicted using both macro-causal graphs and micro-causal graphs, and the root mean square error (RMSE) of the two predictions is calculated. The micro-prediction error is defined as the RMSE of the processing accuracy level predicted using the micro-causal graph. The predicted micro-quantities are converted into levels, i.e., the predicted processing accuracy level, through an aggregation function. The macro-prediction error is defined as the RMSE of the processing accuracy level directly predicted using the macro-causal graph. Both predictions target the same variable, i.e., the processing accuracy level, and can only be compared. The corrected prediction fidelity loss formula is: macro-prediction error minus micro-prediction error multiplied by the square of the abstract granularity parameter, and then the maximum value is taken with zero. This loss ensures that the macro-prediction accuracy is not significantly lower than the micro-prediction accuracy, and the larger λ is, the greater the allowable deviation.

[0084] The second term is the reversibility consistency loss: for the micro-subgraph M in the training set, forward abstraction and inverse refinement are performed respectively through the forward abstraction encoder and the inverse refinement decoder to obtain the reconstructed subgraph M. prime ; Calculate M and M prime The structural differences between them are measured by the edge edit distance, i.e., from M to M prime The minimum number of edge addition and deletion operations required is calculated by dividing the total number of operations required to transform between two subgraphs by the number of edges in M. The reversibility consistency loss is equal to the edge edit distance value. This loss is calculated with the help of the residual representation output by the reversibility residual component. The residual patterns stored in the reversibility residual component help the reverse refinement decoder to more accurately recover the discarded edges.

[0085] The third term is the information bottleneck regularization loss: the information bottleneck regularization loss equals the amount of information loss minus the effective recovery amount, where the effective recovery amount equals the structural similarity between the reconstructed subgraph and the original subgraph multiplied by the amount of information loss; the structural similarity is calculated by comparing the adjacency matrix element by element, that is, the proportion of corresponding elements in the adjacency matrix of the two subgraphs being the same, and the value is between 0 and 1; the role of the information bottleneck regularization loss is to force the system to retain as much transfer-related information as possible during abstraction, and avoid discarding potentially useful details.

[0086] The total loss function is calculated as follows: first, the information bottleneck regularization loss is multiplied by the balance coefficient 0.1, and then the calculation result, prediction fidelity loss, and reversibility consistency loss are added together to obtain the total loss function; the balance coefficient 0.1 is determined through preliminary experiments to make the magnitudes of the three losses comparable.

[0087] Step 3.4: Establish an abstract granularity adaptive cost function and Bayesian optimization.

[0088] The effect feedback of the migration task is defined as a function F of λ; F is equal to the prediction accuracy under the target condition multiplied by 0.5 plus the decrease in physical residual multiplied by 0.5 minus λ multiplied by 0.1 multiplied by the amount of information loss, where the prediction accuracy and the decrease in physical residual are both normalized to values ​​between 0 and 1.

[0089] Since the transfer task effect function F has no direct derivative relationship with the training loss, this method uses Bayesian optimization to update λ: with F as the objective function, after each transfer task, the F value corresponding to the current λ is collected, a Gaussian process surrogate model is constructed, and the point with the greatest expected improvement is selected as the next λ value; after each transfer task is completed, the current abstract granularity parameter and the corresponding effect function value are collected, the Gaussian process surrogate model is updated, and then the abstract granularity parameter used in the next iteration is selected.

[0090] After the update, λ is clipped to the range of 0 to 1; after multiple iterations of the migration task, λ converges to the optimal value, which makes F reach its maximum value.

[0091] Step 4: Causal knowledge transfer and decision fusion for new working conditions: For the current target working condition, use the optimal value of λ obtained in Step 3 to perform knowledge transfer and generate parameter decisions.

[0092] Step 4.1: Perform pre-optimization at the abstract granularity before migration.

[0093] When it is necessary to transfer causal knowledge from the source working condition to a new target working condition, a small amount of sample data from 3 to 5 processing cycles of the target working condition is first collected. This is based on the typical configuration of small-sample transfer learning, which can provide sufficient gradient information and is consistent with the actual situation of quickly acquiring a small amount of data in industrial sites. Each processing cycle completes a five-second multi-sensor signal acquisition according to the rules in step 1.1, consistent with the data acquisition batch in step 1.1. These sample data are input into the processing flow of steps 1 and 2 to obtain the spatiotemporal feature vector of the target working condition. This feature vector is input into a pre-trained regression model. This model collects the working condition feature vector of each task and the optimal abstract granularity parameter value obtained by the final convergence of the task from the completed historical transfer tasks as training samples. The optimal abstract granularity parameter value of each task is obtained after convergence of Bayesian optimization iteration in step 3.4, that is, the λ value that maximizes the effect function F of the transfer task under the task. When the number of accumulated tasks exceeds 10, a random forest or Gaussian process regression model is trained. If there is not enough historical data, the default abstract granularity parameter of 0.5 is used.

[0094] Different migration strategies are executed based on the optimal value of λ: if the optimal value of λ > 0.7, it indicates that the target working condition is biased towards high-level abstraction migration, i.e., macroscopic phenomena dominate, and macroscopic cause-effect graphs are selected as the main migration carrier; if the optimal value of λ < 0.3, it indicates that the target working condition is biased towards low-level abstraction migration, i.e., microscopic details dominate, and microscopic cause-effect graphs are selected as the main migration carrier; if 0.3 ≤ optimal value of λ ≤ 0.7, it is biased towards medium-level abstraction migration, and both granularities are used simultaneously and executed through dual-granularity collaborative migration in step 4.3; the granularity thresholds of 0.7 and 0.3 are based on the following: K-means clustering is performed on the distribution of the optimal value of λ in the completed historical migration tasks, K=3, and the centers of the three clusters are approximately 0.15, 0.50 and 0.85, respectively, and the median values ​​between clusters of 0.3 and 0.7 are used as the boundaries for dividing low, medium and high levels of abstraction; if there is insufficient historical data, the empirical values ​​of 0.3 and 0.7 are used as the default thresholds.

[0095] Step 4.2: Perform reversible refinement expansion and refinement migration.

[0096] If the selected migration vector is a macro-causal graph, perform the following migration sub-steps.

[0097] Step 4.2.1: In the source condition, the micro-subgraph that best matches the features of the target condition is abstracted using the forward abstract encoder in Step 3.2 to obtain high-level concept nodes; simultaneously, the information loss of the forward abstract encoder output and the residual representation vector associated with this abstraction process in the reversibility residual component are recorded; the micro-subgraph is mapped to a fixed-length 128-dimensional embedding vector through an independent graph neural network, and then the cosine similarity is calculated with the 128-dimensional spatiotemporal feature vector of the target condition. The graph neural network structure is the same as the forward abstract encoder in Step 3.2 but does not share parameters, or uses the intermediate layer output of the same encoder; in the micro-causal graph knowledge base of the source condition, the cosine similarity between each micro-subgraph and the feature vector of the target condition is calculated, and the subgraph with the highest similarity is selected as the micro-subgraph of the source condition.

[0098] Step 4.2.2: When migrating to the target working condition, the embedded representation of the high-level concept node is concatenated with the context feature vector of the target working condition, and input into the inverse refinement decoder. The inverse refinement decoder outputs Gaussian distributed parameters, from which a latent vector is sampled, and then reconstructed into a micro subgraph M under the target working condition through the graph decoder. target During the reconstruction process, the inverse refinement decoder also retrieves the residual patterns most similar to the current high-level concept node and context features from the reversible residual component. The retrieval method is to calculate the cosine similarity between the high-level concept node and each residual entry stored in the reversible residual component, and select the top three entries with the highest similarity (>0.7). Their residual vectors are then used as prior inputs to the latent space sampling of the inverse refinement decoder, enabling the inverse refinement decoder to selectively recover those discarded details that may be effective in the target working condition. The cosine similarity threshold of 0.7 was determined experimentally to balance matching accuracy and recall.

[0099] Step 4.2.3: Reconstruct the microscopic subgraph M. target As a causal structure under the target operating condition, it is used for subsequent micro-quantity prediction and process parameter optimization.

[0100] If the selected transfer vector is a micro-causal graph, the micro-subgraph in the source condition that is most similar to the target condition is directly copied as the initial structure, and a small number of target condition samples are used for fine-tuning. The fine-tuning learning rate is set to 0.001 and the number of iterations is 10, based on the standard setting of small sample fine-tuning to prevent catastrophic forgetting.

[0101] Step 4.3: Perform dual-granularity collaborative transfer and hierarchical decision fusion.

[0102] When 0.3 ≤ λ (optimal value) ≤ 0.7, two migration paths are executed simultaneously.

[0103] The first path is a coarse-grained path: Starting from the macro-causal graph, the macro-features of the target working condition are input into the macro-causal graph via the physical information neural network routing in step 2. Causal inference is then performed to predict the probability of macro-processing phenomena occurring. The conditional probability table for each node in the macro-causal graph is learned from historical data of the source working condition through maximum likelihood estimation, or is given by experts. Edge weights are used to adjust the probability bias in the conditional probability table; for example, when the edge weight is large, the conditional probability of the child node is higher when the parent node takes a specific value. Specifically, the posterior probability of each macro-phenomenon node is calculated using a Bayesian network inference algorithm, resulting in the probability distributions of three macro-phenomena: flutter probability values ​​range from 0 to 1; the probability distribution of chip morphology categories, including three categories: strip chips, unit chips, and fragmented chips, with each category having a probability between 0 and 1 and a sum of 1; and the probability distribution of processing accuracy deviation levels, including five levels, with each level having a probability between 0 and 1 and a sum of 1. To unify the output format, the integer encoding of the Bayesian network output is converted into a probability distribution using the Softmax function.

[0104] The second path is the fine-grained path: through the reversible refinement expansion in step 4.2, the microscopic subgraph M under the target working condition is obtained. target Then, starting from micro-variables, causal reasoning is performed using a micro-causal graph to predict micro-process results. The micro-predicted values ​​include three scalars: cutting force amplitude prediction, temperature peak prediction, and vibration amplitude prediction. These three scalars are input to an aggregation function. The aggregation function has three independent output heads: the first output head is a 1-dimensional fully connected layer, outputting chatter probability. A sigmoid activation function is used to compress the output value to between 0 and 1, representing the probability of chatter occurring. The second output head is a 3-dimensional fully connected layer, outputting the probability distribution of chip morphology categories, using a sigmoid activation function... The first output is a Softmax activation function; the second output is a 5-dimensional fully connected layer that outputs the probability distribution of machining accuracy deviation levels, using the Softmax activation function; the inputs of all three outputs share the first two layers of the feature extraction network, which consists of three fully connected layers: the first layer has an input dimension of 3 and an output dimension of 16, activated by a linear rectified function; the second layer has an output dimension of 8, activated by a linear rectified function; then each layer is connected to one of the three outputs; finally, the probability values ​​of three macroscopic phenomena are obtained, with chatter probability as a scalar, and chip morphology category and machining accuracy deviation level as probability distributions.

[0105] The prediction results of the two paths are input into a learnable fusion gating network. First, the coarse-grained path prediction results, the fine-grained path prediction results, and the prediction uncertainty measures of the two paths are concatenated. The uncertainty measures are calculated using the Monte Carlo Dropout method. Therefore, it is necessary to ensure that all networks involved in inference include Dropout layers during training and that Dropout is kept on during prediction. Specifically, in the spatiotemporal graph convolutional network in step 1.5, the super network in step 2.2, the graph neural network in step 3.2, and the aggregation network in this step, a Dropout layer with a dropout rate of 0.2 is added after each fully connected layer.

[0106] Twenty forward propagations were performed, and the variance of the 20 prediction results was calculated as an uncertainty measure. The gating network consisted of two fully connected layers, with the input being a concatenated vector and the output being two original weight values. The network output was converted into a probability distribution using a Softmax function, and the first two values ​​in the probability distribution were taken as the coarse-grained path weight and the fine-grained path weight, respectively, with the sum of the two weights being 1. The final prediction result was equal to the coarse-grained path prediction result multiplied by the coarse-grained path weight, plus the fine-grained path prediction result multiplied by the fine-grained path weight. The fusion gating network was trained on a small amount of target working condition verification data with a learning rate set to 0.001 and 5 training epochs, based on the fact that the gating network has a simple structure and can converge with a small number of iterations, thus avoiding overfitting.

[0107] Step 5: Closed-loop evolution and knowledge distillation based on transfer feedback: The effect feedback after the transfer is performed is backpropagated to the reversible abstraction in Step 3 through a differential path, the abstraction-refinement pair parameters and abstraction granularity parameters are updated, and effective knowledge is distilled into the global knowledge base to achieve continuous evolution.

[0108] Step 5.1: Calculate the prediction error after migration and backpropagate it.

[0109] After completing the migration model run in the target operating condition, a batch of processing data for verification is collected, including 10 processing cycles; the prediction error E is calculated by comparing the prediction results from step 4 with the actual values. target For macroscopic phenomenon prediction, the error is the average of the root mean square errors of the three phenomena. For microscopic quantity prediction, the error is the weighted root mean square error. The weights of each microscopic quantity are set to cutting force 0.4, temperature 0.3, and vibration 0.3, respectively, based on the engineering experience that cutting force has the greatest impact on machining quality, followed by temperature and vibration.

[0110] The prediction error E targetAs a scalar loss, it is backpropagated through the chain rule to all components of the reversible abstraction in step 3, including the spatiotemporal graph convolutional network in step 1.5, the supernetwork in step 2.2, the physical information neural network in step 2.4, the forward abstract encoder, the inverse refinement decoder, the lossy residual components in step 3.2, and the aggregation function and gated network in step 4.3; the specific backpropagation path is: E target The gradients of the fusion-gated network in step 4 are propagated to the two prediction paths; the gradients of the coarse-grained path are propagated to the edge weight parameters of the macro-causal graph and the parameters related to macro-inference in the physical information neural network in step 2.4; the gradients of the fine-grained path are propagated to the decoding parameters of the inverse refinement decoder and the retrieval parameters of the reversibility residual component; then, the gradients are further propagated from the inverse refinement decoder to the encoding parameters of the forward abstract encoder and λ; in addition, the spatiotemporal graph convolutional network in step 1.5 and the supernetwork in step 2.2 also participate in the forward computation, and their parameters are backpropagated from the final error to obtain the gradients through the chain rule; during the gradient propagation process, the learning rate of 0.0005 is determined in the range of 0.0001 to 0.01 through grid search. In the preliminary experiment, this value can make the loss function decrease stably without diverging within 100 iterations. The momentum parameter is set to 0.9, based on the conservative learning rate of deep network backpropagation to prevent gradient explosion.

[0111] Step 5.2: Update all component parameters of the reversible abstraction.

[0112] Based on the gradient calculated in step 5.1, update the parameters of the following components: The weights and biases of the fully connected layers in the forward abstract encoder are updated with a step size equal to the learning rate × gradient. The variational autoencoder parameters of the inverse refinement decoder are included, as well as the network parameters of the encoder mean and variance and the graph convolutional layer parameters of the graph decoder. The entries in the memory matrix of the reversible residual component are updated using an exponential moving average with an update factor of 0.9, meaning the new entry value equals 0.9 multiplied by the old entry value, plus 0.1 multiplied by the new difference map embedding. The abstract granularity parameter λ is updated according to the Bayesian optimization method described in step 3.4. The initial value of λ is set to 0.5 during the first transfer task. After each transfer task, the Gaussian process surrogate model is updated according to the new transfer task effect function value, and the next abstract granularity parameter value is selected. There is no need to set a fixed learning rate.

[0113] After all component parameters are updated, the new total loss is compared with the historical total loss before the update. If the new total loss is at least 5% lower than the old total loss, the update is accepted. The 5% total loss threshold is determined through pre-experiments to ensure sufficient update magnitude while avoiding parameter oscillations. Otherwise, the parameter values ​​before the update are restored, and the learning rate is decayed to 0.9 times the original value, based on the classic adaptive learning rate decay strategy to ensure a steady decline in the optimization process.

[0114] Step 5.3: Perform reversible abstract knowledge distillation and global knowledge base update.

[0115] Distillation is performed on the newly confirmed valid causal knowledge in this transfer task; the distillation process includes three parts: The first part consists of the directed edges in the macro-causal graph that have been verified to be accurate, and the weights of the edges are updated to the average causal effect strength on the validation set. The second part is the newly discovered causal relationships in the micro-causal graph. If the counterfactual intervention test of a certain edge passes the significance test in the target working condition, then the edge and its weight are added to the global micro-causal graph knowledge base. The significance test adopts the t-test, using the change of the outcome variable in all perturbation tests as the sample, with a sample size of at least 30, and calculates the p-value. If the p-value is less than 0.05, the causal relationship is considered to be statistically significant and passes the test. The third part is the abstract mapping relationship, which stores the mapping tuples from the micro subgraph of the source working condition to the macro concept node, as well as the refined mapping tuples verified by the target working condition, into the abstract-refinement global knowledge base.

[0116] The global knowledge base is set to a size of 5,000 entries. This size is based on a margin of approximately twice the number of typical factory processing conditions, ensuring coverage of common conditions while reserving room for expansion. When the size is exceeded, entries are evicted based on their access frequency and most recent usage time. The eviction strategy is least recently used, and separate least recently used queues are maintained for different types of data, such as causal edge queues, equation parameter queues, and abstract mapping queues, to avoid the loss of core knowledge due to mixed eviction.

[0117] Step 5.4: Determine the evolutionary convergence condition and zero-sample transfer capability.

[0118] After multiple cross-condition transfer training sessions, the F-value in step 3.4 is calculated in each transfer task. When the average F-value in five consecutive transfer tasks is greater than the average threshold of 0.85, the standard deviation of the optimal λ-value is less than the standard deviation threshold of 0.05, and the average prediction error of the macro-causal graph and the micro-causal graph on the validation set is less than 0.1, the reversible abstraction is considered to have converged to a stable state. 0.85 and 0.05 are determined through 20 sets of simulation experiments covering typical working conditions such as steady-state cutting, chatter, and chip mutation, indicating that the model performance has reached a high level and is stable. At this point, the model is switched to zero-sample transfer mode: for any new target working condition, there is no need to perform the small-sample pre-optimization in step 4.1, but instead, the model is directly... Using the converged empirical value of the optimal λ value, the migration is performed through steps 4.2 and 4.3. The prediction accuracy in the zero-sample migration mode is expected to be more than 90% of that in the mode with a small number of samples. This percentage is determined based on the statistical benchmark obtained from a large number of previous cross-validation experiments. The 90% percentage is based on the statistical results of 50 previous cross-validation experiments, where the average accuracy of the zero-sample migration is 92% of that of the mode with samples, and the standard deviation is 3%. During the zero-sample migration process, a validation batch still needs to be collected every 20 processing cycles to calculate the prediction error. If the error continues to exceed 95% of that in the mode with samples for 3 consecutive times, the system will automatically revert to the mode with samples for pre-optimization and readjustment with a small number of samples.

[0119] This concludes the entire process of the parameter optimization method for machining the end joint of the internal drive guide rod.

[0120] Example 2: Please see Figure 4 Based on Embodiment 1, this embodiment also provides a parameter optimization system for machining the end joint of the internal drive guide rod, including: The sensor signal preprocessing and dynamic topology construction module is used to resample, align, and differentially filter the cutting signals collected by multiple sensors, generate a theoretical correlation graph based on physical laws and calculate a data-driven correlation matrix, fuse the theoretical correlation graph and the data-driven correlation matrix into a dynamic graph topology matrix based on the confidence score of the previous iteration, and extract spatiotemporal feature vectors based on the dynamic graph topology matrix through spatiotemporal graph convolution. The physical constraint optimization and confidence assessment module is used to construct an equation branch library covering various cutting physical states based on spatiotemporal feature vectors and dynamic graph topology matrices. It generates an equation structure prototype through hypernetwork routing and selects an equation branch as a physical constraint using an exploration and balancing mechanism. During the backpropagation of data fitting loss and physical constraint loss, it performs orthogonal projection protection on conflict gradients. After parameter optimization, it calculates the confidence score and feeds the score back to the sensor signal preprocessing and dynamic topology construction module. The causal knowledge abstraction and granularity optimization module is used to conduct counterfactual intervention tests on microscopic physical variables based on selected physical equation branches to construct microscopic causal graphs, construct macroscopic causal graphs based on expert knowledge, and perform lossy abstraction and refinement processing on microscopic and macroscopic causal graphs through forward abstraction, reverse refinement and residual memory, and obtain the optimal abstract granularity parameters based on Bayesian optimization of abstract granularity parameters. The knowledge transfer and decision fusion module is used to select a transfer strategy based on the value of the optimal abstract granularity parameter, perform lossy refinement expansion or dual-granularity collaborative transfer processing on the target working condition, and weight and fuse the macro causal reasoning results of the coarse-grained path and the micro causal reasoning results of the fine-grained path through a gating network to obtain the comprehensive prediction result under the target working condition. The closed-loop evolution and knowledge distillation module is used to calculate the prediction error and perform backpropagation based on the actual processing results and comprehensive prediction results of the target working condition, update the trainable parameters of the forward link, distill the verified causal knowledge into the global knowledge base, and manage the elimination queues of different data types in the global knowledge base.

[0121] Furthermore, in the sensor signal preprocessing and dynamic topology construction module, the resampling alignment of multi-sensor signals involves uniformly resampling all sensor signals to the same sampling frequency. Differential filtering uses high-pass filtering for cutting force signals, acceleration signals, and acoustic emission signals, while using low-pass filtering or smoothing for temperature signals.

[0122] In the sensor signal preprocessing and dynamic topology construction module, the method of fusing the theoretical correlation graph and the data-driven correlation matrix based on the confidence score is as follows: when the score is greater than the first threshold, the theoretical correlation graph is fully trusted; when the score is less than the second threshold, the data-driven matrix is ​​fully trusted; when the score is between the two thresholds, the score is linearly related to the fusion weight, and the dynamic graph topology matrix is ​​equal to the fusion weight multiplied by the theoretical correlation graph plus one minus the fusion weight multiplied by the data-driven matrix.

[0123] The equation branch library in the Physical Constraint Optimization and Confidence Assessment module covers physical equations under steady-state cutting, chattering, and abrupt chip morphology changes. It explores the use of a balance mechanism to maintain a value function and selects equation branches with a decaying exploration rate. Orthogonal projection protects the system by using only the projected data gradient to update parameters when the data gradient and physical gradient directions conflict.

[0124] The counterfactual intervention test in the causal knowledge abstraction and granularity optimization module perturbs each candidate causal variable with its historical maximum value, historical minimum value, and random offset. The causal relationship is determined based on the change in the predicted value before and after the perturbation. The lossy abstraction refinement process encodes the micro subgraph into a macro embedding through a forward encoder and outputs the amount of information loss. The micro subgraph is reconstructed based on the macro embedding and contextual features through an inverse decoder. The reconstructed difference information is stored through a residual memory component.

[0125] The dual-granularity collaborative transfer in the knowledge transfer and decision fusion module includes coarse-grained paths and fine-grained paths. The coarse-grained path outputs the probability of macroscopic phenomena based on the macro causal graph, while the fine-grained path converts the microscopic quantity prediction into the probability of macroscopic phenomena with three output heads through an aggregation function. The first output head of the aggregation function uses the Sigmoid activation function, and the second and third output heads use the Softmax activation function. The gating network assigns fusion weights based on the prediction uncertainty of the two paths.

[0126] The global knowledge base in the closed-loop evolution and knowledge distillation module maintains independent least recently used elimination queues for causal edges, equation parameters, and abstract mapping relationships. When the average value of the migration effect function and the fluctuation of the optimal abstract granularity parameter of multiple consecutive migration tasks meet the convergence threshold and the prediction error is lower than the accuracy threshold, the mode switches to directly perform knowledge migration without pre-optimization of a small number of samples.

[0127] The system operates as follows: In the machining scenario of the end joint of the internally driven guide rod, the system first acquires cutting signals in real time through force, temperature, vibration, and acoustic emission sensors. The sensor signal preprocessing and dynamic topology construction module performs resampling alignment and differential filtering, and merges the theoretical correlation graph and data-driven correlation matrix into a dynamic graph topology matrix based on the confidence score of the previous iteration. Then, it extracts the spatiotemporal feature vector of the fused physical prior through spatiotemporal graph convolution. The physical constraint optimization and confidence evaluation module selects the physical equation that best matches the current working condition from the equation branch library covering steady state, chatter, and chip abrupt change states as the constraint based on this feature vector. Gradient orthogonal projection is used to protect the physical consistency during parameter updates. After training, the confidence score is output for the next iteration. The iterative fusion weight calculation; the causal knowledge abstraction and granularity optimization module uses the selected physical equations to perform counterfactual intervention tests on micro variables to construct a micro causal graph, and extracts causal knowledge by refining triples through lossy abstraction, and obtains the optimal abstract granularity parameters through Bayesian optimization; the knowledge transfer and decision fusion module selects a transfer strategy based on these parameters, and simultaneously performs coarse-grained macro causal reasoning and fine-grained micro causal reasoning under the medium-grained working condition of easy flutter in thin-walled joints, and outputs the comprehensive prediction result by weighted fusion of the two path results through a gating network; the closed-loop evolution and knowledge distillation module updates the trainable parameters of the entire forward link based on the backpropagation of the error between the actual processing result and the prediction result, and continuously stores the verified causal knowledge into the global knowledge base to realize the continuous evolution of the model across working conditions.

[0128] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0129] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0130] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A parameter optimization method for machining the end joint of an internal drive guide rod, characterized in that, Includes the following steps: By acquiring and preprocessing signals from the cutting process using multiple sensors, generating a theoretical correlation graph based on physical laws, and calculating a data-driven correlation matrix, a dynamic graph topology matrix is ​​obtained through a gated fusion network, and then spatiotemporal feature vectors are extracted. Based on spatiotemporal feature vectors and dynamic graph topology matrices, a physical equation branch library covering steady-state, flutter, and chip abrupt states is constructed. The equation structure prototype is output through a hypernetwork routing generator. Active exploration and utilization balance are introduced to select a physical equation branch. Bidirectional gradient coupling and orthogonal projection protection are performed to optimize parameters and obtain the physical model confidence score. The confidence score is fed back to the gated fusion network of the next iteration. Based on the selected branches of the physical equation, causal inference is performed on microscopic physical variables and macroscopic processing phenomena. Microscopic causal graphs and macroscopic causal graphs are constructed. Causal knowledge is extracted by reversible abstraction and refinement of triples. The optimal abstraction granularity parameters are obtained based on triple information bottleneck loss and Bayesian optimization. Based on the optimal abstract granularity parameter, reversible refinement expansion or dual-granularity collaborative migration processing is performed on the target working condition. The prediction results of coarse-grained path and fine-grained path are weighted and fused through a fusion gating network to obtain the comprehensive prediction result under the target working condition. The prediction error is calculated based on the comprehensive prediction results and the actual processing results of the target working condition. Backpropagation is performed to update the trainable parameters of the forward link, and effective causal knowledge is distilled into the global knowledge base to complete the closed-loop update.

2. The parameter optimization method for machining the end joint of the internal drive guide rod according to claim 1, characterized in that, The process of acquiring and preprocessing signals from the cutting process using multiple sensors includes: acquiring a sensor signal set using multiple source sensors, wherein the sensor signal set includes at least: temperature signal, cutting force signal, acceleration signal and acoustic emission signal; downsampling the sensor signal set to the same frequency; and using different high-pass filter cutoff frequencies for different types of signals, wherein the cutoff frequency of the temperature signal is lower than that of other signals. The process of obtaining the dynamic graph topology matrix through the gated fusion network includes: the gated fusion network calculates scalar fusion weights based on the physical model confidence score obtained in the previous iteration, and performs element-wise weighted summation of the theoretical association graph and the data-driven association matrix to obtain the dynamic graph topology matrix.

3. The parameter optimization method for machining the end joint of the internal drive guide rod according to claim 1, characterized in that, The physical equation branch library includes heat conduction equations and constitutive equations under steady-state cutting conditions, dynamic cutting force modulation equations under chattering conditions, and transient thermal shock equations and dynamic evolution equations of friction coefficients under abrupt changes in chip morphology. The hypernetwork routing generator consists of a fully connected layer, outputs a multidimensional structure vector and divides it into multiple sets of coefficients, which are used to adjust the adjustable parameters of each physical equation branch to generate an equation structure prototype.

4. The parameter optimization method for machining the end joint of the internal drive guide rod according to claim 1, characterized in that, The process of actively exploring and utilizing the balance to select a branch of a physical equation includes: by maintaining the equation structure value function, randomly selecting a branch of the physical equation with an exploration rate or by selecting the maximum value of the value function, and applying a Gaussian perturbation to the adjustable parameter coefficients of the selected branch's equation after each selection; the value function is updated in the form of an exponential moving average and has a minimum lower limit.

5. The parameter optimization method for machining the end joint of the internal drive guide rod according to claim 1, characterized in that, The process of performing parameter optimization by executing bidirectional gradient coupling and orthogonal projection protection includes: calculating the data gradient of the data fitting loss and the physical gradient of the physical constraint loss; if the dot product of the two is negative, then the data gradient is projected onto a subspace orthogonal to the physical gradient, and only the projected data gradient is used to update the network parameters. The confidence score of the physical model is obtained by performing exponential smoothing filtering on the root mean square error of the residuals of consecutive batches, taking the inverse number, and then transforming it through an exponential function.

6. The parameter optimization method for machining the end joint of the internal drive guide rod according to claim 1, characterized in that, The process of causal inference includes: applying historical maximum, historical minimum and random offset perturbations to candidate causal variables, and performing forward inference under physical constraints to determine causal relationships; The reversible abstract refinement triplet includes a forward abstract encoder, a reverse refinement decoder, and a reversibility residual component, used to encode, reconstruct, and store difference information of micro-causal graph subgraphs; The process of obtaining the optimal abstract granularity parameter includes: a forward abstract encoder encodes the subgraph of the micro causal graph to obtain the embedded representation of the macro concept node and outputs the information loss; a reverse refinement decoder reconstructs the micro subgraph based on the embedded representation and context features; a reversible residual component stores the difference information between the reconstructed subgraph and the original subgraph; and Bayesian optimization aims to maximize the transfer effect function and iteratively searches to obtain the optimal abstract granularity parameter.

7. The parameter optimization method for machining the end joint of the internal drive guide rod according to claim 1, characterized in that, The dual-granularity collaborative migration process includes: dividing the migration strategy into three types based on the clustering results of the abstract granularity parameters of historical tasks, including high-level abstraction migration, medium-level abstraction migration, and low-level abstraction migration; the medium-level abstraction migration executes two migration paths simultaneously, including a coarse-grained path and a fine-grained path. The coarse-grained path predicts the probability distribution of macroscopic processing phenomena based on the macroscopic causal graph, while the fine-grained path predicts microscopic process results based on the reconstructed microscopic subgraph and converts them into macroscopic phenomenon probability values. The fine-grained path includes three output heads, which output the chatter probability, chip morphology category probability distribution, and processing accuracy deviation level probability distribution, respectively. The fusion gating network calculates the prediction uncertainty measure of the coarse-grained path and the fine-grained path, and assigns weights to the prediction results of the two paths for weighted summation to obtain the comprehensive prediction result.

8. The parameter optimization method for machining the end joint of the internal drive guide rod according to claim 1, characterized in that, The process of distilling effective causal knowledge into the global knowledge base includes: storing the directed edge weights of the macro-causal graph that have been verified accurately, the new micro-causal relationships that have passed the significance test, and the verified abstract mapping relationships into the global knowledge base; and directly performing knowledge transfer when the average value of the transfer effect function of multiple consecutive transfer tasks is greater than the preset average value threshold and the fluctuation of the optimal abstract granularity parameter is less than the preset standard deviation threshold.

9. A parameter optimization system for machining the end joint of an internal drive guide rod, characterized in that, include: The sensor signal preprocessing and dynamic topology construction module performs resampling alignment and differential filtering on the cutting signals collected by multiple sensors, generates a theoretical correlation graph and calculates a data-driven correlation matrix. Based on the confidence score from the previous iteration, the two are fused to obtain a dynamic graph topology matrix, and spatiotemporal feature vectors are extracted through spatiotemporal graph convolution. The physical constraint optimization and confidence assessment module constructs an equation branch library based on spatiotemporal feature vectors and dynamic graph topology matrix. It generates an equation structure prototype through a hypernetwork and selects an equation branch using an exploration and balancing mechanism. During backpropagation, it performs orthogonal projection protection on conflict gradients. After optimization, it calculates the confidence score and feeds it back to the sensor signal preprocessing and dynamic topology construction module. The causal knowledge abstraction and granularity optimization module constructs a micro causal graph by performing counterfactual intervention tests on microscopic physical variables based on the selected equation branches, and constructs a macro causal graph. It performs lossy abstraction and refinement through forward abstraction, reverse refinement and residual memory, and obtains the optimal abstraction granularity parameters based on Bayesian optimization. The knowledge transfer and decision fusion module selects a transfer strategy based on the optimal abstract granularity parameter, performs lossy refinement expansion or dual-granularity collaborative transfer on the target working condition, and weights and fuses the results of coarse-grained path and fine-grained path through a gating network to obtain a comprehensive prediction result. The closed-loop evolution and knowledge distillation module calculates the prediction error based on the actual processing results and the comprehensive prediction results, backpropagates it, updates the trainable parameters, distills effective causal knowledge into the global knowledge base, and manages the elimination queue of different types of data.

10. The parameter optimization system for machining the end joint of an internal drive guide rod according to claim 9, characterized in that, In the sensor signal preprocessing and dynamic topology construction module, resampling alignment unifies all sensor signals to the same frequency, and differential filtering uses high-pass filtering for cutting force, acceleration and acoustic emission signals, and low-pass filtering for temperature signals. In the physical constraint optimization and confidence assessment module, the equation branch library includes physical equations under steady-state, flutter, and chip mutation states. The exploration uses a balance mechanism to maintain the value function and selects branches with a decaying exploration rate. Orthogonal projection protects the parameters by using only the projected data gradient to update them when there is a gradient conflict. In the causal knowledge abstraction and granularity optimization module, the counterfactual intervention test perturbs each candidate causal variable with its historical maximum value, minimum value, and random offset, and determines the causal relationship based on the amount of change. Lossy abstraction refinement encodes micro-subgraphs into macro-embeddings and outputs the amount of information loss through a forward encoder, reconstructs micro-subgraphs based on macro-embeddings and contextual features through an inverse decoder, and stores the reconstructed difference information through a residual memory component; In the knowledge transfer and decision fusion module, the dual-granularity collaborative transfer includes a coarse-grained path and a fine-grained path. The coarse-grained path outputs the probability of macro phenomena based on the macro causal graph, while the fine-grained path converts the micro quantity prediction into the probability of macro phenomena of three output heads through an aggregation function. The gating network allocates fusion weights based on the prediction uncertainty. In the closed-loop evolution and knowledge distillation module, the global knowledge base maintains independent least recently used elimination queues for causal edges, equation parameters, and abstract mappings. When the average effect function and the fluctuation of the optimal abstract granularity parameter of multiple consecutive migration tasks meet the convergence threshold and the prediction error is lower than the accuracy threshold, the system switches to the direct knowledge migration mode.

Images