A cloud-edge collaboration-based integrated circuit board cutting path dynamic optimization method

Abstract

The application discloses a kind of integrated circuit board cutting path dynamic optimization methods based on cloud edge cooperation, belong to integrated circuit manufacturing technical field, including in the edge execution micro-cutting probe test of board process, collect multi-source dynamic signal, extract cutting response fingerprint by time-frequency domain analysis and upload to cloud end etc. Step.The present application is constructed by process data package based on causal diagram, and the root cause path parameter is identified by causal inversion to defect batch, and the process knowledge rule is generated based on root cause path parameter and is issued to edge end to establish path decision logic, and the path decision logic of multiple edge equipment is uploaded to federal hub for consistency check, conflict resolution and generalization reasoning to form general process knowledge, solve the technical problems that existing technology process knowledge is not interpretable and cannot be inherited across equipment, realize the transformation from black box weight evolution to interpretable knowledge rule and cross-device federal evolution.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit manufacturing technology, specifically relating to a dynamic optimization method for integrated circuit board cutting paths based on cloud-edge collaboration. Background Technology

[0002] Integrated circuit board manufacturing is a core link in the electronic information industry. The quality of PCB cutting directly affects signal integrity, assembly yield, and long-term operational reliability. With the rapid development of 5G communication, new energy vehicles, and high-end consumer electronics, new types of boards such as high-frequency and high-speed boards, thick copper boards, aluminum-based boards, and flexible boards are constantly emerging. Their multi-layer composite structures, precise circuit layouts, and stringent signal integrity requirements have significantly increased the complexity of the cutting process.

[0003] Currently, the industry generally adopts a cloud-edge collaborative architecture to improve the level of manufacturing intelligence, offloading complex computing tasks to the cloud and having the edge device handle execution and simple feedback. However, in the field of PCB cutting path optimization, existing technologies still mainly rely on static physical property perception for path planning, making it difficult to characterize the true dynamic cutting response of materials; path optimization is mostly a single-path post-compensation, unable to compare the advantages and disadvantages of multiple candidate paths from a global perspective before cutting; process evolution relies on black-box model weight updates, resulting in uninterpretable knowledge and isolated devices; and the edge device lacks autonomous decision-making capabilities. These problems restrict the stability of cutting quality and the efficiency of cross-device collaboration. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide a dynamic optimization method for integrated circuit board cutting paths based on cloud-edge collaboration. This method solves the technical problems of existing technologies where process knowledge is uninterpretable and cannot be transferred across devices, achieving a transformation from black-box weighted evolution to interpretable knowledge rules and cross-device federated evolution. The specific solution is as follows: This application provides a method for dynamically optimizing the cutting path of an integrated circuit board based on cloud-edge collaboration, including the following steps: Micro-cutting probe tests were performed at the edge of the panelization process to collect multi-source dynamic signals. Cutting response fingerprints were extracted through time-frequency domain analysis and uploaded to the cloud. The cutting response fingerprint is input into the preset digital twin model to output constitutive parameters through inversion. The digital twin model is calibrated based on the constitutive parameters. A candidate path set is generated based on the calibrated digital twin model and a virtual pre-simulation is performed. A risk heat map is output. A Pareto front is constructed based on the risk heat map and sent to the edge. Based on the preset production task priority, the target path is selected in the Pareto front. Cutting is performed according to the cutting process parameters corresponding to the target path. After cutting, the cut quality index is obtained. The cut quality index is associated with the cutting process parameters and the cutting response fingerprint to form a process data package and upload it to the cloud. A cause-effect graph is constructed in the cloud based on the process data package. Defect batches are identified based on the cut quality index. The cause-effect graph is used to perform causal inversion on the defect batches to identify root cause path parameters. Process knowledge rules are generated based on the root cause path parameters and distributed to the edge. The edge establishes path decision logic based on the process knowledge rules. A federated hub is established in the cloud, and the path decision logic of multiple edge devices is uploaded to the federated hub. The path decision logic is subjected to consistency verification, conflict resolution and generalization reasoning to form general process knowledge. The general process knowledge is then distributed to the edge devices. The edge devices update the path decision logic based on the general process knowledge, reselect the target path based on the updated path decision logic, and output the dynamic optimization result of the cutting path.

[0005] In one embodiment, a micro-cutting probe test is performed at the edge of the panelization process to collect multi-source dynamic signals. The cutting response fingerprint is then extracted through time-frequency domain analysis, including: Select the probe position on the PCB process edge, control the cutting power to a preset ratio of the product cutting power, and perform micro-cutting probe test along a straight line segment of preset length; Simultaneously acquire triaxial cutting force signals, spindle vibration acceleration signals, incision infrared thermal radiation timing signals, and acoustic emission signals; Power spectrum analysis is performed on the force signal to extract the force signal spectral centroid; envelope analysis is performed on the vibration acceleration signal to extract the vibration envelope entropy; exponential decay fitting is performed on the infrared thermal radiation time series signal to extract the thermal radiation decay time constant; and event counting is performed on the acoustic emission signal to extract the acoustic emission impact count rate. The force signal spectrum centroid, vibration envelope entropy, thermal radiation decay time constant, and acoustic emission impact count rate are combined to form a cutting response fingerprint.

[0006] In one embodiment, the digital twin model that receives the cut response fingerprint input is used to invert and output constitutive parameters. The digital twin model is then calibrated based on these constitutive parameters, including: The cutting response fingerprint is input into a preset digital twin model, and the shear modulus, interfacial bonding strength, glass transition temperature and diffusion-induced stress coefficient are output by the digital twin model and used as constitutive parameters. Write the constitutive parameters into the digital twin model, replace the default material parameters built into the digital twin model, and complete the digital twin model calibration.

[0007] In one embodiment, a candidate path set is generated based on the calibrated digital twin model, and a virtual simulation is performed to output a risk heatmap, including: Based on the calibrated digital twin model, a candidate path set is generated, which includes different contour cutting sequences, common edge strategies, corner handling methods, and feed rate curves. A virtual simulation is performed on each candidate path in the candidate path set, and the following outputs are obtained: cut roughness distribution map, heat-affected zone width distribution map, interlayer stratification risk index map, electromagnetic coupling interference intensity map, and residual stress field distribution map, thus obtaining a risk thermal map.

[0008] In one embodiment, constructing a Pareto front based on a risk heatmap includes: The comprehensive quality risk index and cutting efficiency index of each path in the candidate path set are calculated based on the risk heat map. Using the comprehensive quality risk index and the cutting efficiency index as optimization objectives, multi-objective optimization is performed on the candidate path set to obtain the Pareto front.

[0009] In one embodiment, a target path is selected from the Pareto front based on a preset production task priority. Cutting is performed according to the cutting process parameters corresponding to the target path. After cutting, the cut quality index is obtained. The cut quality index is associated with the cutting process parameters and the cutting response fingerprint to form a process data package, including: Select the target path from the Pareto front based on the preset production task priority; The path parameter file corresponding to the target path is used as the cutting process parameters to drive the cutting equipment to perform the actual cutting. After cutting, the cut morphology is scanned to obtain the cut roughness, burr height, heat-affected zone width, dimensional deviation and interlayer delamination length as cut quality indicators. The cut quality indicators are correlated with the cutting process parameters and the cutting response fingerprint to form a process data package.

[0010] In one embodiment, a cause-effect graph is constructed in the cloud based on a process data package. Defective batches are identified based on cut quality indicators. Cause-effect inversion is performed on the defective batches according to the cause-effect graph to identify root cause path parameters, including: Extract the cut quality indicators from the process data package, and determine the defective batches based on the cut quality indicators; Extract the cutting response fingerprint, cutting process parameters, and cut quality indicators from the process data package, and construct a cause-effect graph containing response fingerprint feature nodes, cutting process parameter nodes, and cut quality indicator nodes; Calculate the causal effect from each cutting process parameter node to the cut quality index node; The cutting response fingerprints of defective batches are statistically compared with those of historical normal batches to identify the differences. Based on the differential features, the causal path is traced in the causal graph, and the root cause path parameters leading to the defective batch are inverted.

[0011] In one embodiment, process knowledge rules are generated based on root cause path parameters and distributed to the edge device. The edge device establishes path decision logic based on the process knowledge rules, including: Generate process knowledge rules based on root cause path parameters; Distribute process knowledge rules to the edge; The edge end establishes path decision logic based on process knowledge rules, and calls the path decision logic to correct the path in the next round of cutting.

[0012] In one embodiment, a federated hub is established in the cloud, and the path decision logic of multiple edge devices is uploaded to the federated hub. Consistency verification, conflict resolution, and generalization reasoning are performed on the path decision logic to form general process knowledge, including: Establish a federated hub in the cloud and upload the path decision logic of each edge device to the federated hub; The federal hub performs consistency verification on the path decision logic uploaded by each device, and distinguishes the path decision logic into path decision logic with physical contradictions and path decision logic that passes the consistency verification. Conflict resolution is performed on path decision logic that has physical contradictions, while retaining the path decision logic associated with the root cause path parameters; Generalized reasoning is performed on the path decision logic that has passed the consistency check and the path decision logic that is retained after conflict resolution. The process knowledge rules associated with the root cause path parameters in the path decision logic are extended into general rules that cover a wider range of scenarios, forming general process knowledge. Beneficial effects

[0013] (1) This invention constructs a causal graph based on process data packages, performs causal inversion on defect batches to identify root cause path parameters, generates process knowledge rules based on root cause path parameters and sends them to the edge to establish path decision logic, and uploads the path decision logic of multiple edge devices to the federation hub for consistency verification, conflict resolution and generalization reasoning to form general process knowledge. This solves the technical problem that existing process knowledge is unexplainable and cannot be passed across devices, and realizes the transformation from black box weight evolution to explainable knowledge rules and cross-device federation evolution.

[0014] (2) This invention extracts the cutting response fingerprint by performing micro-cutting probe tests on the side of the panel process, inputs the cutting response fingerprint into the digital twin model to invert and output constitutive parameters and calibrate the digital twin model, generates a candidate path set based on the calibrated digital twin model and performs virtual pre-running to output a risk heat map, and constructs a Pareto front based on the risk heat map. This solves the technical problems of existing technologies being unable to characterize the real dynamic cutting response of materials and unable to perform multi-path global optimization before cutting, and realizes the upgrade from static property perception to dynamic response fingerprint and the leap from single-path post-compensation to multi-path global optimization. (3) This invention updates the path decision logic and reselects the target path based on general process knowledge at the edge end, and outputs the dynamic optimization result of the cutting path, which solves the technical problem that the edge end lacks autonomous decision-making ability in the prior art and realizes the continuous evolution of dynamic optimization of the cutting path. Attached Figure Description

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0016] Figure 1 This is a flowchart provided for the present invention; Figure 2 The flowchart for cutting response fingerprint extraction and uploading provided by this invention; Figure 3 The flowchart for cloud-based twin model calibration and multi-path virtual pre-simulation provided by this invention; Figure 4 The flowchart for edge path decision-making, execution, and quality data acquisition provided by this invention; Figure 5 The causal inversion and knowledge rule evolution flowchart provided by this invention; Figure 6 This is a schematic diagram of the federated knowledge sharing and edge decision logic update provided by the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0018] In related technologies, the following problems typically exist when optimizing the cutting path of integrated circuit boards: The current technology for optimizing the cutting path of PCB boards is based solely on static physical property perception, which cannot characterize the actual dynamic cutting response of the material, resulting in large fluctuations in cutting quality. Existing technologies only perform post-process offset compensation for a single predetermined path, and cannot perform global optimization of multiple paths before cutting, thus missing the optimization space at the path selection level. Existing technologies achieve evolution through weight updates of black-box models, resulting in uninterpretable process knowledge that cannot be passed across devices, and long process debugging cycles. The existing cloud-edge collaborative architecture has a large communication load, lacks autonomous decision-making capabilities at the edge, and has isolated process knowledge among multiple devices, requiring exploration from scratch for new types of boards.

[0019] Therefore, this application provides a dynamic optimization method for integrated circuit board dicing paths based on cloud-edge collaboration, which can provide a dynamic optimization basis for integrated circuit board dicing. Please refer to... Figure 1 , Figure 1 This is a flowchart illustrating the dynamic optimization method for integrated circuit board dicing paths based on cloud-edge collaboration provided in this application embodiment. The dynamic optimization method for integrated circuit board dicing paths based on cloud-edge collaboration provided in this application embodiment is applied in a terminal device, and the method is executed through application software installed in the terminal device. The terminal device can be an edge computing gateway, a cloud server, an industrial control computer, etc.

[0020] It should be noted that the application scenarios described in the following embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0021] The following section provides a detailed description of the dynamic optimization method for integrated circuit board cutting paths based on cloud-edge collaboration provided in this application.

[0022] like Figure 1 As shown, the method includes the following steps S110 to S150.

[0023] S110. Perform micro-cutting probe tests at the edge of the panelization process, collect multi-source dynamic signals, extract cutting response fingerprints through time-frequency domain analysis, and upload them to the cloud.

[0024] Specifically, the micro-cutting probe test is performed on the process edge area around the PCB panel. This process edge is the waste area to be removed before subsequent component placement. Using this area for the micro-cutting probe test will not damage the effective product area. The cutting power of the micro-cutting probe test is set to 10% to 15% of the product's cutting power. This ratio is based on the principle of obtaining effective dynamic signals at the process edge while avoiding damage to the process edge. Verification showed that below 10%, the acoustic emission signal-to-noise ratio is below 20dB, and above 15%, there is a risk of microcracks propagating from the process edge to the effective area. The cutting path is a straight segment with a length set to 1mm to 2mm. This length is based on the principle of covering at least one complete tool rotation cycle and the length of the material's thermal diffusion characteristics. Calibration showed that below 1mm, the thermal radiation signal did not reach a steady state, and above 2mm, the difference from the actual cutting behavior increased.

[0025] In some embodiments, such as Figure 2 As shown, step S110 includes steps S111 to S114.

[0026] S111. Select the probe position at the process edge and perform a micro-cutting probe test along a straight line segment of preset length.

[0027] S112. Synchronously acquire triaxial cutting force signal, spindle vibration acceleration signal, incision infrared thermal radiation timing signal and acoustic emission signal as multi-source dynamic signal.

[0028] S113. Perform power spectrum analysis on the force signal to extract the force signal spectral centroid, perform envelope analysis on the vibration acceleration signal to extract the vibration envelope entropy, perform exponential decay fitting on the infrared thermal radiation time series signal to extract the thermal radiation decay time constant, and perform event counting on the acoustic emission signal to extract the acoustic emission impact count rate.

[0029] Center of gravity of force signal spectrum The calculation formula is: ; In the formula, The centroid of the force signal spectrum, measured in Hz; This represents the i-th frequency component, in Hz. The corresponding power spectral density is expressed in W / Hz; N is the total number of frequency components, which is dimensionless.

[0030] Vibrational envelope entropy The calculation formula is: ; In the formula, The vibrational envelope entropy is dimensionless. Let M be the probability density of the j-th normalized energy interval of the vibration envelope signal, which is dimensionless; M is the number of interval divisions, which is taken as 32, which is also dimensionless. The reason for this value is that the envelope signal has approximately 1024 effective points. After being divided into 32 intervals, each interval has approximately 32 points, which satisfies the statistical stability of the probability density estimation. In addition, 32 is a power of 2, which is compatible with the data alignment of the digital processing system.

[0031] The thermal radiation decay time constant τth is obtained by performing exponential decay fitting on the time-series infrared thermal radiation signal of the notch. The fitting function is: ; In the formula, The surface temperature of the cut at time t is in Kelvin (K). This is the initial temperature rise, expressed in Kelvin (K). is the thermal radiation decay time constant, in seconds; t represents ambient temperature in Kelvin (K); t represents time in seconds (s).

[0032] Acoustic emission impact count rate The calculation formula is: ; In the formula, The acoustic emission impact count rate is expressed in units of 1 / s. The number of acoustic emission impact events during the micro-cutting probe test period is dimensionless. The duration of the experiment is expressed in seconds (s).

[0033] S114. Combine the centroid of the force signal spectrum, the vibration envelope entropy, the thermal radiation decay time constant, and the acoustic emission impact count rate to form a cutting response fingerprint.

[0034] The vector form of the cut response fingerprint F is: ; In the formula, F is the cutting response fingerprint, which is a column vector composed of four feature components; The unit is Hz; Dimensionless; The unit is seconds (s). The unit is 1 / s.

[0035] The cut response fingerprint is compressed into a lightweight feature vector and uploaded to the cloud twin pre-simulation engine via 5G or industrial Ethernet.

[0036] S120. Input the cutting response fingerprint into the preset digital twin model and output constitutive parameters through inversion. Calibrate the digital twin model based on the constitutive parameters. Generate a candidate path set based on the calibrated digital twin model and perform virtual pre-simulation. Output a risk heat map. Construct a Pareto front based on the risk heat map and send it to the edge.

[0037] In some embodiments, such as Figure 3 As shown, step S120 includes steps S121 to S124.

[0038] S121. Input the cutting response fingerprint into the preset digital twin model, and output the shear modulus, interfacial bonding strength, glass transition temperature and diffusion-induced stress coefficient as constitutive parameters by inversion from the digital twin model.

[0039] The digital twin model incorporates a response fingerprint and constitutive parameter mapping network, which is an offline-trained multilayer perceptron. The network has 4 input layer nodes, corresponding to the four feature components of the cut response fingerprint; 6 hidden layers, each with 256 nodes, using a linear rectified function for activation; and 4 output layer nodes, corresponding to shear modulus, interfacial bonding strength, glass transition temperature, and diffusion-induced stress coefficient, using linear activation. The network weights are initialized using the He initialization method, and the biases are initialized to 0. The training dataset includes historical micro-cutting probe experimental data and ground truth constitutive parameters from corresponding destructive cutting verification experiments, with a sample size of at least 5000 sets. The training process uses the Adam optimizer with an initial learning rate of 0.001, 500 iterations, and a batch size of 64. The loss function is the mean squared error function, calculated using the following formula: ; In the formula, The mean squared error loss is dimensionless. To determine the training sample size, take It is 5000; Let be the constitutive parameter truth vector of the k-th sample, where the shear modulus is in MPa, the interfacial bonding strength is in MPa, the glass transition temperature is in K, and the diffusion-induced stress coefficient is dimensionless. This is the constitutive parameter vector for the k-th sample predicted by the network, with the same units as the ground truth vector.

[0040] The digital twin model also includes a viscoelastic constitutive sub-model, a thermo-mechanically coupled damage sub-model, and an interlaminar fracture sub-model. The viscoelastic constitutive sub-model adopts the Prony series form, and the evolution of the shear modulus over time is as follows: ; In the formula, Here is the shear modulus at time t, in MPa; This is the instantaneous shear modulus, expressed in MPa. is the shear modulus coefficient of the i-th Prony series, in MPa; is the relaxation time of the i-th Prony series, in seconds; n is the number of series terms, taken as 3, dimensionless. This number of series terms is based on the fact that the viscoelastic response of the FR4 substrate in the cutting frequency range can be fully characterized by three characteristic relaxation times. Dynamic mechanical analysis has verified that when n is 3, the fitting correlation coefficient is higher than 0.95.

[0041] The thermo-mechanical coupled damage sub-model considers the coupling effect of thermal expansion coefficient and temperature field on stress field. The formula for calculating thermal stress is: ; In the formula, For thermal stress, the unit is MPa; for elastic modulus, the unit is MPa. The coefficient of thermal expansion is 1 / K. Poisson's ratio is dimensionless; T is the current temperature in K. Reference temperature, unit: K; For Kronecker notation, dimensionless.

[0042] The interlayer fracture sub-model adopts the cohesive force model, and the relationship between the interface traction force and the separation displacement is as follows: ; In the formula, This is the normal traction force, expressed in MPa. The interfacial bonding strength is expressed in MPa. This represents the normal separation displacement, in mm. The characteristic separation displacement is expressed in mm. The value is 0.05 mm, which is based on the average characteristic failure displacement of the interface between the copper foil and the FR4 substrate in a standard peel test.

[0043] S122. Write the constitutive parameters into the digital twin model, replace the default material parameters built into the digital twin model, and complete the digital twin model calibration.

[0044] S123. Generate a candidate path set based on the calibrated digital twin model and perform a virtual rehearsal, outputting a risk heat map.

[0045] A candidate path set is generated based on the calibrated digital twin model. The candidate path set contains 8 to 16 candidate paths, covering different contour cutting sequences, edge sharing strategies, corner handling methods, and feed rate curves. Each candidate path in the candidate path set is virtually pre-simulated using finite element discrete element coupled simulation with a time step of 1 μs and a spatial mesh size of 0.05 mm. The simulation duration covers the entire cutting cycle.

[0046] The boundary conditions for the virtual simulation are set as follows: the bottom surface of the workpiece is subject to fixed constraints, limiting the X, Y, and Z-axis displacements; the sides of the workpiece are subject to symmetrical constraints, limiting the normal displacement; the contact between the tool and the workpiece uses a penalty function contact algorithm, with the penalty stiffness set to 10 times the elastic modulus of the softer material. This multiple is based on the accuracy of the equilibrium calculation and the convergence stability. After verification of mesh independence, a penalty stiffness of 10 times can make the contact penetration less than 1% of the mesh size; the tool thermal boundary uses a moving heat source model, and the heat flux density is calculated based on the cutting power and the spot area; the outer surface of the workpiece uses a convective heat transfer boundary, with a convective heat transfer coefficient of 25 W / m². 2K is based on the median value of the natural convection heat transfer coefficient under static air conditions; the interlayer fracture failure criterion adopts the maximum traction force criterion. When the interfacial traction force reaches the interfacial bonding strength, the interfacial unit begins to soften, and the softening process is controlled by the cohesive force model.

[0047] The virtual simulation outputs a risk heat map, which includes a cut roughness distribution map, a heat-affected zone width distribution map, an interlayer stratification risk index map, an electromagnetic coupling interference intensity map, and a residual stress field distribution map.

[0048] In one embodiment, the comprehensive quality risk index and cutting efficiency index of each path in the candidate path set are calculated based on the risk heatmap. The formula for calculating the comprehensive quality risk index vector R is: ; In the formula, R is the comprehensive quality risk index vector, which is a column vector composed of five components; This represents the average roughness of the cut surface, in μm. This represents the maximum width of the heat-affected zone, in μm. The maximum inter-layer stratification risk index is dimensionless. This represents the maximum electromagnetic coupling interference intensity, expressed in V / m. This represents the maximum residual stress, expressed in MPa.

[0049] The formula for calculating the cutting efficiency index E is: ; In the formula, E is the cutting efficiency index, which is dimensionless; Effective cutting time, measured in seconds; The total period time is expressed in seconds.

[0050] S124. Construct a Pareto front based on the risk heatmap and distribute it to the edge.

[0051] Using the comprehensive quality risk index vector and the cutting efficiency index as optimization objectives, a multi-objective optimization is performed on the candidate path set to obtain the Pareto front. The multi-objective optimization employs a pre-defined multi-objective evolutionary algorithm with a population size of 200, a generation count of 100, a crossover probability of 0.9, and a mutation probability of 0.1. The population size of 200 is based on the fact that the candidate path set size is between 8 and 16, and the population size needs to be at least 10 times the number of candidate paths to ensure solution space coverage. Experiments have verified that 200 is a convergent lower bound. The Pareto front consists of a set of non-dominated solutions, including path numbers, corresponding path parameter files, and a summary of the quality risk index. The Pareto front is then distributed to the edge. The obtained set of non-dominated solutions is taken as the Pareto front, denoted as […]. It includes the path number, the corresponding path parameter file, and a summary of quality risk indicators, and distributes the Pareto front to the edge.

[0052] S130. Select the target path in the Pareto front according to the preset production task priority, execute the cutting according to the cutting process parameters corresponding to the target path, obtain the cut quality index after the cutting is completed, associate the cut quality index with the cutting process parameters and the cutting response fingerprint, form a process data package and upload it to the cloud.

[0053] In some embodiments, such as Figure 4 As shown, step S130 includes steps S131 to S134.

[0054] S131. Select the target path from the Pareto front based on the preset production task priority function.

[0055] The preset production task priority function includes quality-first mode, efficiency-first mode, and balanced mode. The quality-first mode selects the path with the smallest norm of the comprehensive quality risk index in the Pareto front, i.e., the target path number. satisfy: ; In the formula, The target path is numbered, dimensionless; P is the Pareto front solution set. Let R be the vector of comprehensive quality risk indicators for the k-th candidate path.

[0056] The efficiency-first approach selects the path with the highest efficiency index in the Pareto front, i.e.: ; In the formula, Let be the cutting efficiency index for the k-th candidate path, which is dimensionless.

[0057] The equilibrium model selects the path closest to the ideal point in the Pareto front. The ideal point is the composite vector of the optimal values ​​of each objective. The distance metric used is Euclidean distance.

[0058] S132. Load the path parameter file corresponding to the target path as the cutting process parameters, and drive the cutting equipment to perform the actual cutting.

[0059] S133. After cutting, scan the cut morphology to obtain the cut roughness, burr height, heat-affected zone width, dimensional deviation and interlayer delamination length as cut quality indicators.

[0060] The roughness of the cut surface was measured by a white light interferometer or a profilometer, in μm; the burr height was measured by an edge extraction algorithm of a visual inspection system, in μm; the width of the heat-affected zone was determined by microhardness testing or metallographic sectioning, in μm; the dimensional deviation was obtained by comparing the visual measurement system with the Gerber design file, in μm; and the interlayer delamination length was obtained by random sampling using an ultrasonic scanning microscope, in μm.

[0061] S134. Associate the cut quality indicators with the cutting process parameters and the cutting response fingerprint to form a process data package and upload it to the cloud.

[0062] The data structure of process data package D is as follows: ; In the formula, D is the process data package; F is the cutting response fingerprint; Number the target path; This is a set of cutting process parameters; This refers to the roughness of the cut surface, expressed in μm. This refers to the burr height, in μm. Width of the heat-affected zone, in μm; This refers to dimensional tolerances, expressed in μm. This represents the interlayer separation length, in μm.

[0063] S140. Construct a cause-effect graph in the cloud based on the process data package, identify defective batches based on the cut quality indicators, perform cause-effect inversion on the defective batches based on the cause-effect graph, identify root cause path parameters, generate process knowledge rules based on the root cause path parameters and distribute them to the edge end, and establish path decision logic on the edge end based on the process knowledge rules.

[0064] In some embodiments, such as Figure 5 As shown, step S140 includes steps S141 to S144.

[0065] S141. Extract the cut quality indicators from the process data package and determine the defective batches based on the cut quality indicators.

[0066] Defective batches are identified by comparing the cut surface quality indicators with preset quality thresholds. The preset quality thresholds are determined based on the IPC-A-600 standard and customer-customized quality specifications. Among these, the cut surface roughness threshold... The value is set at 3.2μm, derived from the surface quality requirements of Level 3 high reliability boards in IPC-A-600; burr height threshold. The value is set at 50μm, derived from the minimum safe clearance requirement between pads and component leads during board assembly; heat-affected zone width threshold. The value is set at 100 μm, derived from experimental statistics of the diffusion distance at the glass transition temperature of the FR4 substrate; interlayer delamination length threshold. The value is set to 0.5mm, which is derived from the maximum allowable layered expansion in IPC-A-600.

[0067] S142. Extract the cutting response fingerprint, cutting process parameters, and cut quality indicators from the process data package, and construct a cause-effect graph containing response fingerprint feature nodes, cutting process parameter nodes, and cut quality indicator nodes.

[0068] The causal graph construction employs a PC algorithm based on conditional independence testing. First, a completely undirected graph is constructed, with the node set including response fingerprint feature nodes, cutting process parameter nodes, and cut quality index nodes. Second, a conditional independence test is performed on each node pair, using the partial correlation coefficient after Fisher-Z transform as the test statistic, with a significance level set to 0.05. This level is based on controlling the Type I error rate within an acceptable range. If a node pair is conditionally independent after controlling for other nodes, the corresponding edge is removed. Third, the remaining edges are oriented according to the collision point identification rule: if node X and node Y have no direct edge but share a common neighbor Z, and Z is not in any separating set of X and Y, then Z is a collision point, and the edge orientation is X→Z←Y. Finally, the Meek rule is applied recursively to the remaining unoriented edges for orientation, resulting in a causal graph in the form of a directed acyclic graph.

[0069] S143. Calculate the causal effect from each cutting process parameter node to the cut quality index node, statistically compare the cutting response fingerprint of the defective batch with the historical normal batch, identify the difference characteristics, trace the causal path in the causal graph based on the difference characteristics, and inversely derive the root cause path parameters that lead to the defective batch.

[0070] The causal effect is measured using the average causal effect, and the calculation formula is as follows: ; In the formula, The average causal effect is dimensionless. For the quality indicators of the cut; This refers to the cutting process parameter node; do indicates the intervention operation; and These represent two different value levels for the cutting process parameters, with the same unit. .

[0071] The average causal effect was specifically estimated using a regression adjustment method based on a causal graph. First, the cutting process parameter nodes were identified based on the causal graph. To the cut quality indicator node First, identify all backdoor paths along the causal path; second, perform regression adjustment on the node set of each backdoor path to establish... right And the multiple linear regression model of the backdoor path nodes; then, fix the backdoor path nodes to their mean, and only change Take values, calculate The predicted values ​​differ; finally, the estimated results of multiple backdoor paths are averaged to obtain a numerical estimate of the average causal effect.

[0072] The differential feature identification uses a two-sample t-test, and the formula for calculating the test statistic t is: ; In the formula, is the test statistic for characteristic f, which is dimensionless; This represents the mean of feature f in the abnormal batches, with units based on feature type; This represents the mean value of feature f in the normal batch, in units of... ; The standard deviation of characteristic f in the abnormal batch is expressed in units of 1 / 2 oz. ; The standard deviation of characteristic f in a normal batch is expressed in units of 1 / 2 oz. ; This refers to the sample size of the abnormal batch, which is dimensionless. This is the normal batch sample size, dimensionless. When... When the absolute value of f is greater than 2.576, it is determined that feature f has a significant difference, and this threshold corresponds to a 99% confidence level.

[0073] S144. Generate process knowledge rules based on root cause path parameters, and distribute the process knowledge rules to the edge end. The edge end establishes path decision logic based on the process knowledge rules.

[0074] The process knowledge rules adopt a condition-action-target triplet form. The condition part is the feature region satisfied by the cutting response fingerprint, the action part is the adjusted cutting process parameters, and the target part is the type of cut quality defect to be avoided. For example, if the centroid of the force signal spectrum is greater than 3.5kHz and the vibration envelope entropy is greater than 2.8, the feed rate will be reduced by 15% and the corner curvature radius will be increased to 0.5mm to avoid excessive cut roughness and burrs. The basis for reducing the feed rate by 15% in this rule is that causal effect analysis shows that for every 10% reduction in feed rate, the average roughness decreases by about 8%, and a 15% reduction can control the roughness below the threshold without significantly sacrificing efficiency; the basis for increasing the corner curvature radius to 0.5mm is that finite element simulation shows that this radius can reduce the stress concentration factor at the corner from 2.5 to 1.8, which is below the stratification risk threshold.

[0075] S150. Establish a federated hub in the cloud, upload the path decision logic of multiple edge devices to the federated hub, perform consistency verification, conflict resolution and generalization reasoning on the path decision logic to form general process knowledge, distribute the general process knowledge to the edge end, update the path decision logic according to the general process knowledge, reselect the target path according to the updated path decision logic, and output the dynamic optimization result of the cutting path.

[0076] In some embodiments, such as Figure 6 As shown, step S150 includes steps S151 to S154.

[0077] S151. Establish a federated hub in the cloud and upload the path decision logic of each edge device to the federated hub.

[0078] S152. Perform consistency verification on the path decision logic, and distinguish the path decision logic into path decision logic with physical contradictions and path decision logic that passes the consistency verification.

[0079] Consistency verification employs logical consistency rules. For any two path decision logics, their condition and action parts are extracted. If the condition parts of the two path decision logics are the same or overlap, while the action parts provide opposite adjustment directions for the same cutting process parameter, then a physical contradiction exists. For example, if the condition of path decision logic A is that the centroid of the force signal spectrum is greater than 3.5kHz, and the action is to decrease the feed rate; and the condition of path decision logic B is that the centroid of the force signal spectrum is greater than 3.5kHz, and the action is to increase the feed rate, then a physical contradiction exists.

[0080] S153. Resolve conflicts in path decision-making logic that has physical contradictions, and retain the path decision-making logic associated with the root cause path parameters.

[0081] Conflict resolution employs a voting mechanism based on the absolute value of the average causal effect. First, the cutting process parameter nodes and cut quality index nodes involved in the path decision logic with physical contradictions are extracted. Second, the absolute value of the average causal effect from the cutting process parameter node to each relevant cut quality index node is calculated. Third, the absolute values ​​of the average causal effect corresponding to the path decision logics of the two conflicting parties are compared, and the one with the larger absolute value is retained. If the difference in the absolute values ​​of the average causal effect is less than a preset tolerance of 0.05, the conflict is considered to be of similar intensity. In this case, the path decision logic with the more conservative adjustment range is retained, i.e., the one with the smaller parameter adjustment range is prioritized.

[0082] S154. Perform generalized reasoning on the path decision logic that has passed the consistency check and the path decision logic that is retained after conflict resolution. Extend the process knowledge rules associated with the root cause path parameters in the path decision logic into general rules that cover a wider range of scenarios to form general process knowledge.

[0083] Generalized inference employs inductive logic programming to extend specific numerical conditions into general rules. For example, the rule that a 15% reduction in feed rate occurs when the centroid of the force signal spectrum is greater than 3.5kHz is generalized to a 10% to 20% reduction in feed rate when the centroid of the force signal spectrum is in the high-frequency region. Here, the high-frequency region is defined as the upper 20% quantile of the historical data distribution where the centroid of the force signal spectrum is located, and the 10% to 20% reduction range is based on the statistical confidence interval of effective reductions verified under similar conditions in the historical rule base.

[0084] The edge device updates its path decision logic based on general process knowledge. When processing a new type of sheet material for the first time or encountering abnormal working conditions not covered by history, it sends a knowledge query request containing the current cutting response fingerprint to the federated center. Based on the similarity matching of the cutting response fingerprint, the federated center retrieves the historical process knowledge rules of the most similar scenario from the general process knowledge. The retrieved historical process knowledge rules are used as the initial strategy recommendation and sent to the edge device. The edge device updates its path decision logic based on the initial strategy recommendation, reselects the target path according to the updated path decision logic, and outputs the dynamic optimization result of the cutting path.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic optimization method for integrated circuit board cutting paths based on cloud-edge collaboration, characterized in that, Includes the following steps: Micro-cutting probe tests were performed at the edge of the panelization process to collect multi-source dynamic signals. Cutting response fingerprints were extracted through time-frequency domain analysis and uploaded to the cloud. The cutting response fingerprint is input into the preset digital twin model to output constitutive parameters through inversion. The digital twin model is calibrated based on the constitutive parameters. A candidate path set is generated based on the calibrated digital twin model and a virtual pre-simulation is performed. A risk heat map is output. A Pareto front is constructed based on the risk heat map and sent to the edge. Based on the preset production task priority, the target path is selected in the Pareto front. Cutting is performed according to the cutting process parameters corresponding to the target path. After cutting, the cut quality index is obtained. The cut quality index is associated with the cutting process parameters and the cutting response fingerprint to form a process data package and upload it to the cloud. A cause-effect graph is constructed in the cloud based on the process data package. Defect batches are identified based on the cut quality index. The cause-effect graph is used to perform causal inversion on the defect batches to identify root cause path parameters. Process knowledge rules are generated based on the root cause path parameters and distributed to the edge. The edge establishes path decision logic based on the process knowledge rules. A federated hub is established in the cloud, and the path decision logic of multiple edge devices is uploaded to the federated hub. The path decision logic is subjected to consistency verification, conflict resolution and generalization reasoning to form general process knowledge. The general process knowledge is then distributed to the edge devices. The edge devices update the path decision logic based on the general process knowledge, reselect the target path based on the updated path decision logic, and output the dynamic optimization result of the cutting path.

2. The method for dynamic optimization of integrated circuit board cutting path based on cloud-edge collaboration according to claim 1, characterized in that, Micro-cutting probe tests were performed at the edge of the panelization process to collect multi-source dynamic signals. Cutting response fingerprints were extracted through time-frequency domain analysis, including: Select the probe position on the PCB process edge, control the cutting power to a preset ratio of the product cutting power, and perform micro-cutting probe test along a straight line segment of preset length; Simultaneously acquire triaxial cutting force signals, spindle vibration acceleration signals, incision infrared thermal radiation timing signals, and acoustic emission signals; Power spectrum analysis is performed on the force signal to extract the force signal spectral centroid; envelope analysis is performed on the vibration acceleration signal to extract the vibration envelope entropy; exponential decay fitting is performed on the infrared thermal radiation time series signal to extract the thermal radiation decay time constant; and event counting is performed on the acoustic emission signal to extract the acoustic emission impact count rate. The force signal spectrum centroid, vibration envelope entropy, thermal radiation decay time constant, and acoustic emission impact count rate are combined to form a cutting response fingerprint.

3. The method for dynamic optimization of integrated circuit board cutting path based on cloud-edge collaboration according to claim 1, characterized in that, The digital twin model is input into the cut response fingerprint and inverted to output constitutive parameters. The digital twin model is then calibrated based on these constitutive parameters, including: The cutting response fingerprint is input into a preset digital twin model, and the shear modulus, interfacial bonding strength, glass transition temperature and diffusion-induced stress coefficient are output by the digital twin model and used as constitutive parameters. Write the constitutive parameters into the digital twin model, replace the default material parameters built into the digital twin model, and complete the digital twin model calibration.

4. The method for dynamic optimization of integrated circuit board cutting path based on cloud-edge collaboration according to claim 1, characterized in that, Based on the calibrated digital twin model, a candidate path set is generated and a virtual simulation is performed, outputting a risk heatmap, including: Based on the calibrated digital twin model, a candidate path set is generated, which includes different contour cutting sequences, common edge strategies, corner handling methods, and feed rate curves. A virtual simulation is performed on each candidate path in the candidate path set, and the following outputs are obtained: cut roughness distribution map, heat-affected zone width distribution map, interlayer stratification risk index map, electromagnetic coupling interference intensity map, and residual stress field distribution map, thus obtaining a risk thermal map.

5. The method for dynamic optimization of integrated circuit board cutting path based on cloud-edge collaboration according to claim 1, characterized in that, Constructing the Pareto front based on risk heatmaps, including: The comprehensive quality risk index and cutting efficiency index of each path in the candidate path set are calculated based on the risk heat map. Using the comprehensive quality risk index and the cutting efficiency index as optimization objectives, multi-objective optimization is performed on the candidate path set to obtain the Pareto front.

6. The method for dynamic optimization of integrated circuit board cutting path based on cloud-edge collaboration according to claim 1, characterized in that, Based on the preset production task priority, a target path is selected from the Pareto front. Cutting is performed according to the cutting process parameters corresponding to the target path. After cutting, the cut quality index is obtained. The cut quality index is then associated with the cutting process parameters and the cutting response fingerprint to form a process data package, including: Select the target path from the Pareto front based on the preset production task priority; The path parameter file corresponding to the target path is used as the cutting process parameters to drive the cutting equipment to perform the actual cutting. After cutting, the cut morphology is scanned to obtain the cut roughness, burr height, heat-affected zone width, dimensional deviation and interlayer delamination length as cut quality indicators. The cut quality indicators are correlated with the cutting process parameters and the cutting response fingerprint to form a process data package.

7. The method for dynamic optimization of integrated circuit board cutting path based on cloud-edge collaboration according to claim 1, characterized in that, A cause-effect graph is constructed in the cloud based on the process data package. Defective batches are identified based on the cut quality indicators. The cause-effect graph is used to perform causal inversion on the defective batches to identify root cause path parameters, including: Extract the cut quality indicators from the process data package, and determine the defective batches based on the cut quality indicators; Extract the cutting response fingerprint, cutting process parameters, and cut quality indicators from the process data package, and construct a cause-effect graph containing response fingerprint feature nodes, cutting process parameter nodes, and cut quality indicator nodes; Calculate the causal effect from each cutting process parameter node to the cut quality index node; The cutting response fingerprints of defective batches are statistically compared with those of historical normal batches to identify the differences. Based on the differential features, the causal path is traced in the causal graph, and the root cause path parameters leading to the defective batch are inverted.

8. The method for dynamic optimization of integrated circuit board cutting path based on cloud-edge collaboration according to claim 1, characterized in that, Process knowledge rules are generated based on root cause path parameters and distributed to the edge computing endpoint. The edge computing endpoint then establishes path decision logic based on these process knowledge rules, including: Generate process knowledge rules based on root cause path parameters; Distribute process knowledge rules to the edge; The edge end establishes path decision logic based on process knowledge rules, and calls the path decision logic to correct the path in the next round of cutting.

9. The method for dynamic optimization of integrated circuit board cutting path based on cloud-edge collaboration according to claim 1, characterized in that, Establish a federated hub in the cloud, upload the path decision logic of multiple edge devices to the federated hub, perform consistency verification, conflict resolution and generalization reasoning on the path decision logic to form general process knowledge, including: establish a federated hub in the cloud, upload the path decision logic of each edge device to the federated hub; The federal hub performs consistency verification on the path decision logic uploaded by each device, and distinguishes the path decision logic into path decision logic with physical contradictions and path decision logic that passes the consistency verification. Conflict resolution is performed on path decision logic that has physical contradictions, while retaining the path decision logic associated with the root cause path parameters; Generalized reasoning is performed on the path decision logic that has passed the consistency check and the path decision logic that is retained after conflict resolution. The process knowledge rules associated with the root cause path parameters in the path decision logic are extended into general rules that cover a wider range of scenarios, forming general process knowledge.

Images