Dynamic knowledge graph and deep learning fused pulse laser regulation and control system

Abstract

The invention discloses a dynamic knowledge graph and deep learning fused pulse laser regulation and control system, which relates to the technical field of pulse laser processing control and comprises a dynamic knowledge graph construction module, a deep learning hybrid model module and a double-closed-loop collaborative optimization module which are connected in sequence. According to the dynamic knowledge graph and deep learning fused pulse laser regulation and control system, through deep cooperation of the dynamic knowledge graph and the deep learning model, accurate analysis and real-time optimization of a pulse laser multi-domain parameter coupling relationship are realized. Through a cross-domain coupling quantization technology of a dynamic knowledge graph, the interactive influence of time domain, space domain and frequency domain parameters is converted into a coupling coefficient matrix which can be updated online, and the problem that the strong coupling of the parameters is difficult to dynamically analyze by a traditional method is solved; through a working condition triggering type calculation architecture of a hybrid model, a sub-module is dynamically activated based on an LSTM-CNN-GNN heterogeneous network, and the calculation load is remarkably reduced while the prediction precision is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of pulsed laser processing control technology, and specifically to a pulsed laser control system that integrates dynamic knowledge graphs and deep learning. Background Art

[0002] In the field of pulsed laser control, the core challenge faced by existing technologies is the strong coupling of time, space and frequency domain parameters in the pulsed laser system, and the cross-domain influence is difficult to dynamically analyze. Specifically, parameters such as the pulse width and repetition frequency in the time domain, the spot size and working distance in the space domain, and the laser wavelength in the frequency domain do not act independently, but are interrelated and influence each other, forming a complex coupling relationship. This complex coupling relationship makes it difficult for traditional control methods to fully and accurately grasp the impact of each parameter on the system performance, and it is impossible to achieve coordinated optimization and control of the multi-domain parameters of the pulsed laser system, which in turn makes it difficult for the technical indicators of the pulsed laser system to meet the diverse needs of high-precision processing scenarios. For example, complex parameter coupling will lead to instability in performance indicators such as processing accuracy, ablation depth, and surface quality, and it is impossible to dynamically adjust and optimize according to actual processing requirements. Therefore, how to effectively solve the problem of complex coupling of time, space and frequency domain parameters of the pulsed laser system and the difficulty in dynamically analyzing cross-domain influences has become the key to achieving multi-dimensional coordinated optimization and control of the pulsed laser system and improving the system's adaptability to high-precision processing scenarios. Summary of the Invention

[0003] To achieve the above objectives, the present invention is implemented through the following technical solutions: a pulsed laser control system integrating dynamic knowledge graph and deep learning, comprising:

[0004] Dynamic knowledge graph construction module, deep learning hybrid model module and dual closed-loop collaborative optimization module;

[0005] The input end of the dynamic knowledge graph construction module is connected to the time domain sensor, spatial domain sensor and frequency domain sensor of the pulse laser system, which is used to collect pulse width, repetition frequency, spot size, working distance and laser wavelength parameters in real time;

[0006] The input end of the deep learning hybrid model module is connected to the output end of the dynamic knowledge graph construction module, which is used to process multi-domain parameters and output predicted values ​​of machining accuracy, ablation depth and surface quality;

[0007] The input end of the dual closed-loop collaborative optimization module is connected to the output end of the deep learning hybrid model module, the control end thereof is connected to the actuator of the pulse laser system, and the feedback end thereof is connected to the output end of the sensor.

[0008] Preferably, the dynamic knowledge graph construction module includes:

[0009] An entity relationship definition unit associates time domain parameters with pulse width and repetition frequency entities, spatial domain parameters with spot size and working distance entities, and frequency domain parameters with laser wavelength entities;

[0010] Cross-domain coupling strength quantification unit generates dynamic coupling coefficient matrices between time domain and space domain, time domain and frequency domain, and space domain and frequency domain;

[0011] The real-time update unit dynamically adjusts the entity node relationship and coupling coefficient matrix according to the sensor data.

[0012] Preferably, the cross-domain coupling strength quantization unit operates in the following manner:

[0013] Initialize the coupling coefficient matrix based on the pulse laser transmission equation;

[0014] The coupling coefficient matrix is ​​updated through training of historical data collected by the photodetector and the spot analyzer;

[0015] When abnormal machining accuracy is detected, real-time correction of the coupling coefficient matrix is ​​triggered.

[0016] Preferably, the deep learning hybrid model module comprises parallel connections of:

[0017] The time series processing unit uses a long short-term memory network with a temporal convolution kernel to process the sequence characteristics of pulse width and repetition frequency;

[0018] The spatial processing unit uses a convolutional neural network with a spatial transformation layer to process the spatial distribution characteristics of the spot size and working distance;

[0019] The frequency domain processing unit uses a graph neural network to embed the feature vectors of laser wavelength nodes in the knowledge graph.

[0020] Preferably, the output end of the deep learning hybrid model module is connected to a dual verification unit, which includes:

[0021] Knowledge graph rule validator, which verifies whether the parameter combination violates the entity relationship constraints;

[0022] Adversarial robustness tester that generates perturbations in the neighborhood of optimized parameters and detects performance fluctuations.

[0023] Preferably, the dual verification unit performs:

[0024] If a parameter combination is judged invalid by the knowledge graph rule validator, the combination is discarded;

[0025] If the adversarial robustness tester detects that the performance fluctuation exceeds the threshold, it triggers model weight fine-tuning.

[0026] Preferably, the dual closed-loop collaborative optimization module includes:

[0027] Parameter optimizer, which uses evolutionary algorithms to search for the optimal parameter combination that meets the target performance;

[0028] An execution controller outputs the verified parameter combination to an actuator of the pulse laser system;

[0029] The incremental learner updates the knowledge graph coupling coefficient matrix and model weights based on the real-time performance data fed back by the sensors.

[0030] Preferably, the incremental learner operates as follows:

[0031] When the deviation between the actual machining accuracy and the model prediction value exceeds the limit, the coupling coefficient matrix of the knowledge graph is updated first;

[0032] If the deviation persists, an incremental learning algorithm is used to adjust the weights of the fully connected layers of the deep learning hybrid model.

[0033] Preferably, it also includes:

[0034] The fault diagnosis unit, whose input is connected to the dynamic knowledge graph construction module, locates the cross-domain coupling fault source by backtracing the entity relationship path when the processing accuracy is abnormal.

[0035] A pulse laser control method for a pulse laser control system based on the fusion of dynamic knowledge graph and deep learning, comprising the following steps:

[0036] Step a: Collect time domain, spatial domain, and frequency domain parameters in real time through sensors;

[0037] Step b: Dynamically update the entity relationships and coupling coefficient matrix of the knowledge graph;

[0038] Step c: predict the system performance of the parameter combination through deep learning hybrid model;

[0039] Step d: Outputting the optimal parameters to the pulse laser system after double verification;

[0040] Step e: Update the knowledge graph and model weights based on real-time performance feedback.

[0041] The present invention provides a pulsed laser control system that integrates dynamic knowledge graphs with deep learning. It has the following beneficial effects:

[0042] This pulsed laser control system, which integrates dynamic knowledge graphs with deep learning, achieves precise analysis and real-time optimization of the coupling relationship between multi-domain parameters of pulsed lasers through the deep collaboration between dynamic knowledge graphs and deep learning models. Through the cross-domain coupling quantification technology of the dynamic knowledge graph, the interactive effects of time, space, and frequency domain parameters are converted into a coupling coefficient matrix that can be updated online, solving the problem that traditional methods have difficulty in dynamically analyzing strong coupling of parameters. Through the working condition-triggered computing architecture of the hybrid model, submodules are dynamically activated based on the LSTM-CNN-GNN heterogeneous network, significantly reducing the computational load while ensuring prediction accuracy. Based on a dual closed-loop verification and incremental learning system, physical conflict parameters are intercepted through knowledge graph rule verification and adversarial robustness testing, and an error-driven hierarchical update strategy is combined to ensure continuous system optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a data flow diagram between modules of the pulse laser control system of the present invention;

[0044] Figure 2 This is a schematic diagram of constructing a cross-entity relationship tracing tree in the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] See also Figure 1 and Figure 2 The present invention provides a technical solution: a pulse laser control system integrating dynamic knowledge graph and deep learning, comprising:

[0047] Dynamic knowledge graph construction module, deep learning hybrid model module and dual closed-loop collaborative optimization module;

[0048] The input end of the dynamic knowledge graph construction module is connected to the time domain sensor, spatial domain sensor and frequency domain sensor of the pulse laser system to collect pulse width, repetition frequency, spot size, working distance and laser wavelength parameters in real time;

[0049] The input end of the deep learning hybrid model module is connected to the output end of the dynamic knowledge graph construction module, which is used to process multi-domain parameters and output the predicted values ​​of machining accuracy, ablation depth and surface quality;

[0050] The input end of the dual closed-loop collaborative optimization module is connected to the output end of the deep learning hybrid model module, its control end is connected to the actuator of the pulse laser system, and its feedback end is connected to the output end of the sensor.

[0051] It should be further explained that during the specific implementation process, when the system starts up, the time domain sensor collects pulse width and repetition frequency in real time, the spatial domain sensor collects spot size and working distance, and the frequency domain sensor collects laser wavelength parameters. The time domain sensor can use a high-speed photodetector, and the spatial domain sensor can use a spot analyzer. After receiving the above parameters, the dynamic knowledge graph construction module performs the following operations:

[0052] S1: Dynamic binding of entity relationships: bind pulse width and repetition frequency as time domain entities, spot size and working distance as space domain entities, and laser wavelength as frequency domain entity;

[0053] Establish a cross-domain coupling relationship library. For example, if the pulse width is detected to increase and the spot size is detected to decrease, an energy density abnormality warning is triggered; if the laser wavelength deviates from the material absorption peak range, the material applicability conflict is marked.

[0054] S2: Deep learning hybrid model collaborative processing: The timing processing unit analyzes the pulse width sequence through the LSTM network. If it detects that the fluctuation of the interval between consecutive pulses exceeds the preset tolerance, it activates the pulse cluster effect compensation submodule;

[0055] The spatial processing unit uses CNN to identify the deformation characteristics of the light spot. If the working distance exceeds the depth of field, resulting in excessive light spot ellipticity, the spatial transformation layer is called to correct the energy distribution.

[0056] The frequency domain processing unit inputs the feature vector of the wavelength node in the knowledge graph into the GNN, and generates a frequency domain risk identifier when the wavelength is in the high-risk range of the material reflection spectrum.

[0057] S3: Dual closed-loop optimization and verification execution: After the parameter optimizer generates candidate parameter combinations, it enters a dual verification channel, including the following channels:

[0058] Rule verification channel: If the wavelength in the combination is equal to 1064nm and the processing object is a polymer material, and it violates the "metal processing applicability" rule of the knowledge graph, it will be discarded directly;

[0059] Robustness test channel: Add ±5% perturbation to the candidate parameter neighborhood to generate test samples. If the fluctuation range of the surface quality prediction value exceeds the safety threshold, the model fine-tuning is initiated.

[0060] The verified parameters are output to the laser system actuator, and the photodetector monitors the actual ablation depth in real time. If the deviation between the actual value and the predicted value exceeds the limit for three consecutive times, the pulse width-spot size coupling coefficient in the knowledge graph is updated first. If the deviation still does not converge after the update, the incremental learner is triggered to adjust the weights of the LSTM-CNN-GNN fusion layer.

[0061] S4: Fault tracing linkage mechanism: When the processing accuracy is abnormal, the fault diagnosis unit backtracks along the knowledge graph relationship edge: If the surface quality degradation is associated with the spatial domain node, and the frequency domain wavelength node confidence is detected to be less than 0.9 at the same time, it is determined to be a cross-domain coupling mismatch fault; if the energy density anomaly is only associated with the time domain-spatial domain edge, it is located as a single-domain parameter actuator drift fault.

[0062] Dynamic knowledge graph building blocks include:

[0063] An entity relationship definition unit associates time domain parameters with pulse width and repetition frequency entities, spatial domain parameters with spot size and working distance entities, and frequency domain parameters with laser wavelength entities;

[0064] Cross-domain coupling strength quantification unit generates dynamic coupling coefficient matrices between time domain and space domain, time domain and frequency domain, and space domain and frequency domain;

[0065] The real-time update unit dynamically adjusts the entity node relationship and coupling coefficient matrix according to the sensor data.

[0066] It should be further explained that in the specific implementation process, after the sensor data is input into the dynamic knowledge graph construction module, the entity relationship definition unit first performs parameter entity processing: if the variance of the pulse width sequence is detected to exceed the time domain stability threshold, the current pulse width entity is marked as a high fluctuation state, and its collaborative constraint relationship with the repetition frequency entity is activated; when the spot ellipticity returned by the spot analyzer is greater than the deformation tolerance, the spot size entity is automatically associated with the spatial distortion correction category, and its topological association weight with the working distance entity is strengthened;

[0067] If the laser wavelength falls into the high-risk range of the material reflection spectrum, such as the high reflection band of polymer materials near 1064nm, the absorption risk indicator of the frequency domain entity is triggered and associated with the material applicability performance node.

[0068] The cross-domain coupling intensity quantization unit synchronously starts the dynamic coupling mechanism: in the initialization stage, the pulse transmission characteristics are derived based on the nonlinear Schrödinger equation to generate the time-domain-space domain basic coupling matrix. For example, when the working distance increases and the spot size expands, the energy density attenuation compensation coefficient is automatically calculated; if the pulse width is shortened and the wavelength is blue-shifted, the frequency domain dispersion compensation factor is generated; in the operation stage, the matrix is ​​corrected in real time through historical data training: when the actual ablation depth deviates from the predicted value for three consecutive times and the error increases in the same direction, the intensive training of the pulse width-ablation depth coupling weight is started; if the surface quality anomaly only occurs under specific wavelength conditions, the confidence score of the wavelength-surface quality relationship edge is updated in a targeted manner.

[0069] The real-time update unit dynamically adjusts the graph based on the sensor stream data, including the following modes:

[0070] Regular update mode: After completing each single processing task, the entity node attributes are refreshed with the average parameter value;

[0071] Abnormal trigger mode: If the processing accuracy drops suddenly beyond the safety threshold, the system will immediately retrace the last 10 parameter combinations. If a sudden change in spot size is detected as the main cause, the weight of the spot size-processing accuracy relationship edge will be reset. If there is also a wavelength offset, the time-domain-frequency domain cross-coupling matrix will be corrected synchronously.

[0072] When the material replacement signal is input: the original time domain-space domain coupling relationship is retained, and the mapping relationship library between the wavelength entity and the new material absorption spectrum is rebuilt.

[0073] Accurate relationship binding is achieved through a dynamic classification mechanism of entity states with high volatility, distortion and risk identification; the limitations of empirical formulas are overcome through a dual-source evolutionary strategy of coupling matrices with physical basis and data-driven correction; and efficiency and real-time requirements are balanced through a three-level update response system of routine, abnormal and emergency, which is significantly different from the construction method of static knowledge graphs.

[0074] The cross-domain coupling strength quantization unit operates in the following way:

[0075] Initialize the coupling coefficient matrix based on the pulse laser transmission equation;

[0076] The coupling coefficient matrix is ​​updated through training of historical data collected by the photodetector and the spot analyzer;

[0077] When abnormal machining accuracy is detected, real-time correction of the coupling coefficient matrix is ​​triggered.

[0078] It should be further explained that during the specific implementation, at the system initialization stage, the cross-domain coupling intensity quantization unit extracts the basic physical constraints from the pulse laser transmission equation:

[0079] When it is detected that the spot size expands due to an increase in the working distance, an energy density attenuation compensation coefficient is automatically generated based on the beam propagation model. This coefficient is nonlinearly positively correlated with the distance increment.

[0080] If the pulse width is shortened to the femtosecond level and accompanied by a wavelength blue shift, the calculation of the frequency domain dispersion compensation factor is triggered. The compensation factor value is determined by the wavelength offset and the material group velocity dispersion characteristics.

[0081] For certain materials, such as ceramics, when the repetition frequency exceeds the critical value of thermal diffusion, a heat accumulation suppression coefficient is generated and bound to the time-domain-surface quality relationship edge.

[0082] When the system is running, the coupling coefficient matrix is ​​dynamically modified through real-time data flow, and the process is as follows:

[0083] Continuous error trigger mechanism: If the actual ablation depth exceeds the predicted value three times in a row and the error increases, the pulse width-ablation depth coupling weight is determined to be too weak, and intensive training of this weight is initiated: nearly 50 sets of high-energy density operating data are collected; the coupling curve is refitted while maintaining the wavelength-material constraints;

[0084] If surface quality anomalies only occur when processing polymers at a wavelength of 1064nm, the wavelength-surface quality relationship edge is locked as the correction target: the initial confidence level of this relationship edge is attenuated to below 0.7; and a polymer material reflectivity compensation parameter is added.

[0085] Multi-condition adaptive strategy: In the micro-hole processing scenario, when the aperture aspect ratio exceeds 5:1, the pulse width-ablation depth coupling weight is strengthened to weaken the influence of the working distance-processing efficiency relationship; if the system switches to the brittle material etching mode, the heat accumulation suppression coefficient is disabled and the wavelength-thermal stress coupling channel is activated.

[0086] Emergency response under abnormal working conditions includes: when irreversible ablation damage occurs at the processing position: trace back the spatial parameters in the last five parameter combinations; if the spot size fluctuation exceeds the safety tolerance and coincides with the damage position in time and space, reset the spot size-processing accuracy relationship edge to the initial weight, triggering the recalibration of the spatial sensor; if the damage is accompanied by a sudden jump in wavelength, such as a jump of ±10nm, freeze the time domain-frequency domain cross-coupling matrix and start the frequency domain stabilization device intervention process.

[0087] Basic reliability is ensured through the priority design of physical constraints and data correction, and precise control is achieved through adaptive coupling channel switching in processing scenarios such as deep holes, brittleness, and differentiated polymer processing. A fault defense system is constructed through the abnormal graded response mechanism from parameter training to equipment intervention, which solves the static nature and scenario blind spot problems of cross-domain coupling analysis in traditional methods.

[0088] The Deep Learning Hybrid Model module contains parallel connections of:

[0089] The time series processing unit uses a long short-term memory network with a temporal convolution kernel to process the sequence characteristics of pulse width and repetition frequency;

[0090] The spatial processing unit uses a convolutional neural network with a spatial transformation layer to process the spatial distribution characteristics of the spot size and working distance;

[0091] The frequency domain processing unit uses a graph neural network to embed the feature vectors of laser wavelength nodes in the knowledge graph.

[0092] It should be further explained that, in the specific implementation process, after the sensor data is input into the deep learning hybrid model module, the timing processing unit first analyzes the pulse width sequence: when it is detected that the coefficient of variation of the continuous pulse interval exceeds the dynamic stability threshold, the time convolution kernel is automatically activated to capture the pulse cluster effect. If the width range within the same pulse cluster exceeds the preset tolerance, the pulse energy balance compensation submodule is triggered; for the repetitive frequency mutation condition, such as switching from 1kHz to 100kHz, the gated memory reset function of LSTM is enabled to clear the historical state to avoid frequency aliasing errors.

[0093] The spatial processing unit synchronously processes the spot features: if the ellipticity index returned by the spot analyzer exceeds the deformation threshold of 0.25, the spatial transformation layer is activated for real-time correction: when the working distance exceeds the depth of field range and causes a ring-shaped spot, an ellipse-to-circle transformation matrix is ​​generated; if the position offset of the reading head causes the center of the spot to drift, the translation compensation vector is calculated; in micro-structure processing, when the feature size is smaller than the spot diameter, the multi-scale convolution kernel group is activated to synchronously extract the local energy distribution and global morphological features.

[0094] The frequency domain processing unit performs deep coupling analysis: the feature vector of the wavelength node in the knowledge graph is input into the GNN. When the wavelength falls into the high-risk range of material reflection, such as 1064nm processing polycarbonate, an absorption risk identifier is generated and passed to the output layer of the hybrid model. If the material is marked as high reflectivity in the knowledge graph, an additional reflection compensation weight factor is added.

[0095] When the system switches to a new wavelength, such as from 1064nm to 532nm: freeze the historical characteristics of the original wavelength node; reconstruct the graph relationship edge between the frequency domain entity and the new material absorption spectrum.

[0096] The feature fusion stage adopts a hierarchical splicing strategy, which includes the following:

[0097] Primary fusion: The pulse cluster feature vector output by LSTM and the spot shape vector extracted by CNN are spliced ​​according to the channel dimension;

[0098] Cross-domain enhancement: When high-frequency pulses and tiny light spots are detected simultaneously, the spatiotemporal feature cross-attention layer is activated;

[0099] Frequency domain intervention: If the absorption risk identifier confidence is > 0.9, reflection compensation scaling is applied to the fused features.

[0100] Efficient processing is achieved through a dynamic allocation mechanism of computing resources, the problem of micro-machining precision is overcome through time-space feature cross-enhancement technology, and the risk of machining failure is prevented through material reflectivity feedforward compensation, which is effectively different from the neural network model with a fixed structure.

[0101] The output of the deep learning hybrid model module is connected to a dual verification unit, which includes:

[0102] Knowledge graph rule validator, which verifies whether the parameter combination violates the entity relationship constraints;

[0103] Adversarial robustness tester that generates perturbations in the neighborhood of optimized parameters and detects performance fluctuations.

[0104] It should be further explained that in the specific implementation process, after the deep learning hybrid model outputs a candidate parameter combination, the knowledge graph rule verifier first performs a physical compliance review: If the wavelength parameter is detected to be at 1064nm and the processing object knowledge graph node is marked as the "polymer material" category, the material absorption conflict rule is triggered, the combination is automatically discarded, and a wavelength tuning suggestion is generated;

[0105] When the pulse width is less than 100fs and the working distance exceeds the depth of field limit, the beam transmission stability rule is activated, and it is judged as a combination of spatial-temporal coupling failure. For micro-hole processing scenarios, if the spot diameter is larger than the target aperture and the multi-beam strategy is not enabled, the spatial resolution constraint rule is violated, and the parameter flow to the execution end is forcibly terminated.

[0106] The parameters that pass the rule verification are fed into the robustness tester to perform stability evaluation, including:

[0107] Perturbation sample generation stage: Generate ±5% random perturbations in the candidate parameter neighborhood. If the pulse width in the original combination is at the femtosecond level, the perturbation range is reduced to ±2%.

[0108] When the spot size approaches the diffraction limit, lateral position perturbation is disabled and only the energy intensity is adjusted;

[0109] Fluctuation detection phase: The disturbance sample is input into the hybrid model to obtain performance prediction values. If the extreme difference of the surface quality index exceeds the upper tolerance limit of the material roughness, it is marked as a highly sensitive combination. When the ablation depth prediction value shows non-monotonic changes, such as fluctuations exceeding three oscillation cycles, it is determined that there is a risk of hidden instability.

[0110] Response decision stage: Parameter solidification protocol is initiated for highly sensitive combinations: core parameters remain unchanged, and only auxiliary variables are optimized. Core parameters include wavelength and pulse width, while auxiliary variables include repetition rate and working distance.

[0111] If latent instability is detected, the adversarial training mode of the hybrid model is immediately triggered: a noise-tolerant regularization term is injected into the hidden layer and iterated until the fluctuation converges.

[0112] The exception handling linkage mechanism includes: when a single batch of parameters fails rule verification three times in a row: backtracking the confidence level of the relevant entity relationship in the knowledge graph; if the confidence level of the "wavelength-material applicability" edge is less than 0.8, initiating the frequency domain entity library reconstruction process;

[0113] When adversarial testing reveals systemic sensitivity: freeze the current optimizer output channel and activate prior knowledge to guide optimization: match similar operating condition parameter templates from historical successful cases.

[0114] The risk of parameter failure is eliminated through the coordinated verification of physical rules and random disturbances. The specific protection of ultra-precision working conditions ensures reliability in extreme scenarios. The multi-level emergency degradation mechanism ensures the continuous availability of the system, solving the problem of processing accidents caused by missing verification in the optimization plan.

[0115] Double verification unit execution:

[0116] If a parameter combination is judged invalid by the knowledge graph rule validator, the combination is discarded;

[0117] If the adversarial robustness tester detects that the performance fluctuation exceeds the threshold, it triggers model weight fine-tuning.

[0118] It should be further explained that, in the specific implementation process, when a candidate parameter combination enters the dual verification unit, the system first performs an in-depth review of the knowledge graph rule verifier: If the wavelength parameter is detected to be at 1064nm and the processing object is marked as a "polymer material" category by the knowledge graph, the material absorption conflict rule is immediately triggered, the combination is discarded, and the wavelength tuning recommended range is fed back to the optimizer; when the pulse width is lower than the critical femtosecond value and the working distance exceeds the depth of field limit of the optical system, the beam transmission instability rule is activated, and it is determined that the combination has the risk of spatial-temporal coupling failure and an alarm code is generated;

[0119] For combinations where the spot diameter is larger than the current processing feature size, if the multi-beam strategy node is not associated, the spatial resolution constraint rule is violated, the execution link is forcibly interrupted, and the self-test process of the micro-optical extension module is activated.

[0120] The parameters that pass the rule verification enter the adversarial robustness tester to perform a multi-level stability evaluation, including the following stages:

[0121] Intelligent perturbation injection stage: Automatically compresses the perturbation amplitude under femtosecond pulse conditions to avoid distortion of ultrashort pulse characteristics; locks the spatial coordinate perturbation channel when the spot approaches the diffraction limit, leaving only the energy intensity adjustment dimension open; and disables random perturbations of thermally sensitive parameters, including repetition frequency, for brittle material processing scenarios.

[0122] Dynamic fluctuation monitoring stage: Real-time tracking of the extreme changes in surface quality prediction values ​​in disturbed samples. If the value exceeds the tolerance limit of the material roughness, it will be marked as a highly sensitive combination. When the ablation depth prediction curve shows non-monotonic oscillation characteristics, the implicit instability analysis algorithm is activated to identify potential risk points.

[0123] Hierarchical decision execution phase: Parameter fixation protocols are enabled for highly sensitive combinations: the core parameters of wavelength and pulse width remain unchanged, and the repetition frequency and working distance are locally optimized through evolutionary algorithms; if oscillatory instability is detected, noise tolerance regularization constraints are immediately injected into the fully connected layer of the hybrid model, and adversarial training is initiated until the output fluctuations converge to within a safe threshold.

[0124] Exception handling uses a multi-level linkage response, including: when continuous rule verification fails, reverse tracing the confidence score of the relationship edge in the knowledge graph, and when the "wavelength-material applicability" edge score continues to fall below the dynamic threshold, triggering the emergency reconstruction process of the frequency domain entity library;

[0125] When adversarial testing finds systemic sensitivity, the real-time optimizer output channel is suspended, the historical case matching engine is activated, and the optimal parameter template is retrieved according to the processing material and precision requirements to implement degradation control.

[0126] Precise protection is achieved through a strong binding mechanism between physical rules and operating conditions, multi-channel disturbance constraint technology prevents failures in extreme scenarios, and an instability risk classification and disposal system ensures continuous and reliable operation, fundamentally breaking through the limitations of single threshold judgment.

[0127] The dual closed-loop collaborative optimization module includes:

[0128] Parameter optimizer, which uses evolutionary algorithms to search for the optimal parameter combination that meets the target performance;

[0129] An execution controller outputs the verified parameter combination to an actuator of the pulse laser system;

[0130] The incremental learner updates the knowledge graph coupling coefficient matrix and model weights based on the real-time performance data fed back by the sensors.

[0131] It should be further explained that, in the specific implementation process, when the parameter combination that has passed the double verification is input into the dual closed-loop collaborative optimization module, the parameter optimizer starts the intelligent search process: in the deep hole processing scenario, if it is detected that the target ablation depth exceeds the critical value of the material thermal diffusion, the parameter sampling is performed preferentially in the low repetition frequency area to avoid the heat accumulation effect causing carbonization of the hole wall; for the brittle material etching task, when the surface quality requirement reaches the submicron level, the pulse width search range is automatically constrained to above the picosecond level to suppress the risk of microcracks caused by stress waves; if the current material is marked as a high reflectivity category in the knowledge graph, the wavelength-reflectivity compensation weight factor is injected into the evolutionary algorithm to guide the population to evolve towards the near-infrared band.

[0132] When the execution controller outputs the optimized parameters to the laser system, monitoring is initiated simultaneously: the actual ablation depth is captured in real time through a high-speed photodetector. If there is an increasing deviation in the same direction during three consecutive processing cycles, the repetition frequency is dynamically adjusted to compensate for the energy gap if the source of the deviation is related to the time domain parameters. If the spatial sensor reports uneven spot energy distribution, the working distance fine-tuning mechanism is immediately involved to refocus.

[0133] In femtosecond micromachining, if the actual surface roughness increases suddenly but the model prediction value is stable, the current parameter execution channel is frozen and the online spectrum analyzer is activated to scan the material phase change characteristics to verify whether amorphization damage has occurred.

[0134] The incremental learner initiates hierarchical updates based on the monitoring error, including the following responses:

[0135] Primary response: When the actual ablation depth deviates from the predicted value three times in a row and the absolute value of the error exceeds the material single-pulse removal threshold, online training of the coupling coefficient of the pulse width-ablation depth relationship edge in the knowledge graph is triggered first, and a sliding window is used to update the fitting curve of the most recent 50 sets of data. If the error reduction in the next cycle after the update is less than 50%, it is marked as a model failure risk.

[0136] Advanced response: For working conditions marked as model failure, the current sensor data stream is extracted to generate an incremental data set, the CNN convolution kernel and LSTM gating weights are kept unchanged, and only the cross parameters of the fully connected layer are fine-tuned. When high-frequency parameters switch scenarios, the historical state protection mechanism is enabled to prevent feature drift.

[0137] Precise avoidance is achieved through the strong binding of material properties and optimized search, hidden failures are blocked by multi-dimensional diagnosis of the execution process, and learning efficiency is guaranteed by intelligent sorting of incremental updates, thus solving the technical defect of the separation of optimization-execution-learning links.

[0138] The incremental learner works as follows:

[0139] When the deviation between the actual machining accuracy and the model prediction value exceeds the limit, the coupling coefficient matrix of the knowledge graph is updated first;

[0140] If the deviation persists, an incremental learning algorithm is used to adjust the weights of the fully connected layers of the deep learning hybrid model.

[0141] It should be further explained that in the specific implementation process, when the actual processing accuracy fed back by the sensor deviates from the model prediction value, the system initiates the hierarchical response protocol of the incremental learner, which includes the following two stages:

[0142] Primary response stage: If the ablation depth deviates in the same direction for three consecutive processing cycles and the absolute value of the deviation exceeds the material single pulse removal threshold, the dynamic update of the pulse width-ablation depth relationship edge in the knowledge graph is triggered first: the latest 50 sets of high-energy density working condition data are collected, and the coupling curve is refitted while maintaining the wavelength-material constraint. When the spot size fluctuation is detected as the main cause, the weight coefficient of the spot size-energy density relationship edge is synchronously corrected. If the deviation reduction in the next processing cycle does not reach the expected improvement level, it is marked as a model failure risk condition.

[0143] Advanced response phase: Initiate neural network fine-tuning for conditions marked as model failures: Lock the CNN convolution kernel and LSTM gating weights to remain constant, and only open the cross-parameters of the fully connected layer for incremental learning; use a small sample sliding window mechanism to extract the current sensor data stream to generate an incremental training set;

[0144] When encountering high-frequency parameter switching scenarios, historical state snapshot protection is enabled, and the mean of the pulse cluster characteristics before the switch is saved as a benchmark reference; if the pulse width distribution deviation at the new frequency exceeds the stability tolerance, the snapshot data is injected to constrain the weight update direction;

[0145] Emergency avoidance mechanism: When systemic instability is detected during incremental learning, if the surface roughness prediction value fluctuates in opposite directions twice in a row, such as first increasing, then decreasing, and then increasing again, the learning process is immediately suspended;

[0146] Activate cross-domain coupling retrospective diagnosis: Check whether the time-domain-spatial domain cross matrix in the knowledge graph is outdated; if the spot analyzer shows ring distortion but the graph is not marked, reset the spatial entity association rules; after confirming that the fault has been resolved, continue incremental learning from the breakpoint and compress the learning rate to a safe level.

[0147] Accurate response is achieved through update priority decisions driven by error characteristics, and state anchoring technology for high-frequency switching scenarios ensures system stability. The full-process monitoring mechanism of incremental learning eliminates the risk of learning out of control, and overcomes the problems of "blind full-weight update" and "lack of scenario adaptability" in incremental learning.

[0148] It also includes a fault diagnosis unit, whose input is connected to the dynamic knowledge graph construction module. When the machining accuracy is abnormal, it locates the source of the cross-domain coupling fault by backtracing the entity relationship path. It should be further explained that in the specific implementation process, when the machining accuracy abnormality triggers the fault diagnosis unit, the system activates the multi-level tracing engine to accurately locate the source of the fault, including the following:

[0149] Primary diagnostic layer: If the surface roughness suddenly increases and the airspace sensor detects that the spot ellipticity exceeds the standard, the associated path of the spot size node in the knowledge graph is immediately traced back. If there is a recent jump record of the working distance, it is determined to be a spot distortion fault caused by defocus, triggering the intervention of the autofocus mechanism. If the readhead position confidence level is continuously lower than 0.8, it is located as a mechanical displacement deviation fault, and the optical calibration protocol is activated.

[0150] Intermediate diagnostic layer: When the ablation depth anomaly is accompanied by a frequency domain wavelength shift: the wavelength-material absorption relationship edge is retrieved along the knowledge graph. If the current wavelength is within the material reflection peak range, an insufficient absorption warning is generated and a switch to the ultraviolet band is recommended. The time-domain-frequency domain cross-coupling matrix is ​​simultaneously checked. If the pulse width is compressed to the femtosecond level but dispersion compensation is not activated, it is marked as a group velocity dispersion mismatch fault and the compensation parameters are injected into the actuator.

[0151] Advanced diagnostic layer: For multi-domain coupling anomalies, such as a sudden drop in processing efficiency accompanied by an expansion of the heat-affected zone: a cross-entity relationship traceability tree is constructed, and high-frequency association paths are detected first; if the end node of the traceability tree contains "material thermal diffusion coefficient exceeds the threshold", the cooling system enhancement instruction is activated.

[0152] A dynamic confidence control mechanism runs through the entire diagnostic process: when a relationship edge misses a real fault three times in a row, its confidence score is automatically attenuated: if the "pulse width-processing efficiency" edge score falls below 0.7, data-driven retraining of the edge is triggered; the diagnostic path is downgraded for edges with scores below 0.5 to avoid misleading subsequent analysis; after successfully locating the fault, the confidence of the related relationship edge is improved: for example, when defocus distortion is correctly identified, the "working distance-spot deformation" edge score is increased by 0.15; if the cross-domain coupling diagnosis time is lower than the preset standard, an additional 0.1 score reward is added.

[0153] Material adaptability verification serves as the ultimate guarantee: When a fault cannot be resolved through parameter adjustment, the micro-area spectral data of the current material surface is extracted and compared with the standard absorption spectrum features in the knowledge graph: if a surge in reflectivity of an unknown absorption peak is detected, a new material sub-graph is created; an emergency processing parameter template is initiated: high-energy pulse blasting mode is enabled for metal impurities, and long pulses with low thermal effect are switched to for organic pollutants.

[0154] Accurate fault location is achieved through a three-level penetrating diagnostic architecture. The dynamic evolution system of confidence through misdiagnosis attenuation, success rewards, and low-score demotion ensures continuous optimization of the map. Emergency response to unknown materials solves the problem of sudden processing failures and addresses the defects of single-level analysis blind spots and static knowledge base rigidity.

[0155] A pulse laser control method for a pulse laser control system based on the fusion of dynamic knowledge graph and deep learning, comprising the following steps:

[0156] Step a: Collect time domain, spatial domain, and frequency domain parameters in real time through sensors;

[0157] Step b: Dynamically update the entity relationships and coupling coefficient matrix of the knowledge graph;

[0158] Step c: predict the system performance of the parameter combination through deep learning hybrid model;

[0159] Step d: Outputting the optimal parameters to the pulse laser system after double verification;

[0160] Step e: Update the knowledge graph and model weights based on real-time performance feedback.

[0161] It should be further explained that, in the specific implementation process, after the system is started, the time domain sensor captures the pulse width sequence in real time. When it is detected that the coefficient of variation of the continuous pulse interval exceeds the dynamic stability threshold, the time convolution kernel is automatically activated to capture the pulse cluster effect; the spatial domain sensor synchronously collects the spot distribution data. If the ellipticity index exceeds the deformation tolerance, the spatial transformation layer is triggered to correct the energy distribution; the wavelength parameters obtained by the frequency domain sensor are input into the knowledge graph, and an absorption risk identifier is generated when the wavelength falls into the high-risk range of material reflection.

[0162] The dynamic knowledge graph construction module receives multi-domain parameters and performs intelligent updates: if the variance of the pulse width sequence continues to exceed the standard, the time domain entity is marked as a high-fluctuation state and its collaborative constraint with the repetition frequency is strengthened; when the working distance jump causes spot distortion, the association weight between the spot size node and the processing accuracy is corrected; when a material replacement signal is detected, the mapping relationship library between the wavelength entity and the absorption spectrum of the new material is reconstructed.

[0163] Dynamic computational optimization is implemented when processing fusion features using the deep learning hybrid model: the CNN spatial transformation layer is enabled only when the spot ellipticity is greater than 0.25 to avoid full-time computing and resource consumption; the spatiotemporal feature cross-attention layer is activated in scenarios where high-frequency pulses and micro-spots coexist, strengthening the weights of key energy density areas; and when the confidence level of the frequency domain risk identifier is greater than 0.9, reflection compensation scaling is applied to the output prediction value.

[0164] Parameter combinations undergo rigorous screening through a dual verification process. Combinations that violate wavelength-material compatibility rules are immediately blocked. Parameters that pass these verifications undergo intelligent perturbation testing, compressing the perturbation amplitude to a safe range for femtosecond pulses and disabling repetition rate perturbations for brittle materials. The validated optimal parameters are then output to the laser system actuator.

[0165] Multi-dimensional monitoring is implemented during the execution process: When the actual ablation depth deviates from the predicted value in the same direction for three consecutive cycles, the repetition frequency is dynamically adjusted to compensate for the energy gap. When the surface roughness increases suddenly but the model prediction is stable, online spectral analysis is initiated to verify the material phase change damage. Real-time errors trigger incremental learning hierarchical responses: the primary response updates the knowledge graph coupling coefficient; the advanced response fine-tunes the model's fully connected layer weights. Historical state snapshots are enabled during high-frequency parameter switching to prevent feature drift.

[0166] Multi-level fault tracing is initiated when machining accuracy is abnormal: spot distortion accompanied by a jump in working distance is identified as a defocus fault; when ablation anomalies coexist with wavelength shift, material absorption spectrum conflicts are identified; and a cross-entity traceability tree is constructed for systematic failure scenarios to locate the root cause of heat accumulation. Each successful diagnosis increases the confidence level of the relevant edge, while repeated misdiagnoses trigger data-driven retraining. The final material verification phase compares micro-area spectral features and, upon identifying unknown impurities, switches to high-energy blasting or cryogenic removal emergency templates.

[0167] System startup and real-time sensing: The time-domain sensor continuously monitors changes in the pulse width sequence. When it detects that the fluctuation amplitude of the interval between consecutive pulses exceeds the set limit for dynamic stability, the system automatically activates the temporal convolution kernel function to capture the pulse cluster effect. The spatial domain sensor simultaneously collects the spot energy distribution data. If the analysis results show that the spot ellipticity exceeds the preset deformation tolerance, the spatial transformation layer is immediately triggered to make real-time corrections to the energy distribution. After the laser wavelength parameters obtained by the frequency domain sensor are input into the knowledge graph, the system automatically compares them to the material absorption characteristics database. If the current wavelength is identified as being in the material's high reflectivity risk range, a frequency domain absorption risk identifier is generated.

[0168] Dynamic construction of knowledge graph: The dynamic knowledge graph construction module performs intelligent updates after receiving multi-domain sensor data: If the fluctuation variance of the pulse width sequence continues to exceed the time domain stability threshold, the system will mark the time domain entity as a high-fluctuation state, and at the same time strengthen its collaborative constraint relationship with the repetition frequency entity; when the working distance parameter mutates and causes spot distortion, the correlation weight coefficient between the spot size node and the processing accuracy performance is automatically corrected; when receiving a material replacement command signal, the system immediately rebuilds the wavelength entity and the absorption spectrum mapping relationship library of the new material, retaining the original entity topology structure.

[0169] Deep learning hybrid model optimization operation: Dynamic resource allocation is implemented when the hybrid model processes features: the spatial transformation layer of the convolutional neural network is enabled only when the spot ellipticity exceeds the deformation threshold to avoid resource consumption of full-time operation; when high-frequency pulse parameters and small spot size parameters coexist, the time domain and spatial domain feature cross-attention layer is automatically activated to enhance the weight focus on key energy density areas; if the confidence score of the frequency domain absorption risk identifier exceeds the dynamic threshold, the system applies a reflection compensation scaling factor to the output prediction value.

[0170] A dual parameter verification mechanism: Candidate parameter combinations are first validated against knowledge graph rules. Violations of the physical rules governing wavelength and material compatibility, such as those for processing polymer materials at near-infrared wavelengths, are immediately intercepted and provided with tuning recommendations. Passing these verifications then enter robustness testing, where the system intelligently adjusts perturbation strategies based on the operating conditions. For femtosecond pulse processing, parameter perturbations are reduced to a safe range, while random perturbations of thermally sensitive parameters are prohibited in brittle material processing scenarios. The validated optimal parameters are then output to the laser actuator.

[0171] Multi-dimensional monitoring of the execution process: During actual machining, if the ablation depth deviates from the predicted value in the same direction for three consecutive machining cycles, the system dynamically adjusts the repetition frequency parameter to compensate for the energy gap. If the measured surface roughness value suddenly deteriorates but the model prediction value remains stable, the online spectral analysis module is immediately activated to scan the phase change characteristics of the material surface to verify whether amorphization damage has occurred. Real-time error data triggers an incremental learning hierarchical response mechanism: the primary response prioritizes updating the coupling coefficients of key relationship edges in the knowledge graph, such as the correlation function between pulse width and ablation depth; the advanced response fine-tunes the weight parameters of the model's fully connected layer, and uses historical state snapshot anchoring technology to prevent feature drift in high-frequency parameter switching scenarios.

[0172] Multi-level fault diagnosis traceability: A hierarchical diagnosis engine is activated when machining accuracy is abnormal. When spot distortion is accompanied by a jump in the working distance, the system identifies an optical defocus fault and triggers the autofocus mechanism. If an ablation depth abnormality is detected and a wavelength shift is detected, the knowledge graph is retrieved for conflicting material absorption spectrum nodes. A cross-entity relationship traceability tree is constructed for systemic failure scenarios, and the source of heat accumulation is located by analyzing terminal nodes such as the thermal diffusion coefficient. The confidence score of the relevant relationship edge is increased after each successful diagnosis, and continuous diagnosis failures trigger a data-driven retraining mechanism. The ultimate guarantee phase identifies unknown impurity components through micro-area spectral feature comparison, switching between high-energy pulse blasting and low-temperature long pulse removal emergency templates based on the impurity type.

[0173] Dynamic Confidence Evolution System: Confidence scores for knowledge graph edges are dynamically managed. If a specific edge fails to locate a true fault during diagnosis three times in a row, the system automatically lowers its confidence score. If the score falls below a set threshold, the edge is downgraded to prevent misleading subsequent analysis. Successful fault location significantly increases the score of related edges, and additional scoring rewards are awarded when cross-domain coupled diagnostic efficiency exceeds a benchmark. Confidence changes are synchronously fed back to the frequency domain processing unit of the hybrid model, enabling the co-evolution of the knowledge base and learning model.

[0174] Emergency Response and System Recovery: If continuous negative fluctuations in the output prediction value are detected during incremental learning, the learning process is immediately paused and cross-domain coupled retrospective diagnostics are initiated. The timeliness of the spatial and temporal cross-matrix in the knowledge graph is checked. If the spot analyzer displays annular distortion but the graph does not update the markers, the spatial entity association rules are reset. After confirming that the fault has been resolved, the learning process is resumed at a compressed learning rate. If an unknown absorption peak is detected during material compatibility verification, a temporary material sub-graph is created and the impurity characteristic spectral fingerprint is recorded. This allows for automatic matching of high-energy blasting or cryogenic removal templates in subsequent processing tasks.

[0175] It should be further explained that during the specific implementation process:

[0176] Knowledge graph update priority: Use sliding window mean update after routine tasks; backtrack multiple groups of parameter combinations to implement cross-validation when accuracy drops suddenly; retain the basic relationship between time domain and spatial domain when materials are replaced to reconstruct frequency domain mapping.

[0177] Hybrid model feature fusion: temporal features and spatial features are spliced ​​according to the channel dimension; when high-frequency micromachining conditions are detected, the cross-attention layer is activated to assign regional weights; the frequency domain risk factor acts as a scaling factor on the final output.

[0178] Incremental learning safety mechanism: Verify feature distribution offset before updating model weights; inject regularization constraints when oscillation characteristics are detected; call historical state snapshots to stabilize the feature space in high-frequency switching scenarios.

[0179] Fault tracing tree construction: Taking the abnormal performance node as the root, search along the relationship edges in descending order of confidence; prioritize the detection of recently changed parameter paths; and use the thermodynamic node as the final inspection filter for systematic failures.

[0180] A pulsed laser control method integrating dynamic knowledge graph and deep learning includes the following steps:

[0181] Step S1: The time domain sensor collects the pulse width sequence in real time. When it detects that the fluctuation of the continuous pulse interval exceeds the dynamic stability threshold, it automatically activates the time convolution kernel to capture the pulse cluster effect; the spatial domain sensor synchronously monitors the spot distribution. If the ellipticity exceeds the deformation tolerance, the spatial transformation layer is triggered to correct the energy distribution; the frequency domain sensor obtains the wavelength parameters and inputs them into the knowledge graph. When the high reflection risk of the material is identified, an absorption risk identifier is generated.

[0182] Step S2: The dynamic knowledge graph construction module performs intelligent updates: when the pulse width fluctuation continues to exceed the standard, the time domain entity is marked as a high-fluctuation state and the collaborative constraint with the repetition frequency is strengthened; when the working distance jump causes spot distortion, the association weight between the spot size node and the processing accuracy is corrected; when the material is replaced, the wavelength entity and the absorption spectrum mapping library of the new material are rebuilt.

[0183] Step S3: The deep learning hybrid model starts dynamic computational optimization: the spatial transformation layer is enabled only when the spot ellipticity exceeds the standard; the scene where high-frequency pulses and micro-spots coexist activates the time-domain and spatial-domain feature cross-attention layer to enhance the focus on key areas of energy density; when the confidence of the frequency-domain risk identifier exceeds the threshold, reflection compensation scaling is applied to the output prediction value.

[0184] Step S4: The candidate parameter combinations enter a dual verification channel: combinations that violate the wavelength material applicability rules are intercepted first; the parameters that pass the verification are subjected to intelligent perturbation testing, namely: the perturbation amplitude is compressed in the femtosecond pulse working condition, and thermally sensitive parameter perturbations are disabled in the brittle material scenario.

[0185] Step S5: The verified optimal parameters are output to the laser actuator, and multi-dimensional monitoring is started at the same time: when the ablation depth deviates from the predicted value continuously and in the same direction, the repetition frequency compensation energy is dynamically adjusted; when the surface roughness suddenly increases but the model prediction is stable, online spectral analysis is started to verify the material phase change damage.

[0186] Step S6: Real-time error triggers incremental learning hierarchical response: the primary response updates the key coupling coefficients of the knowledge graph; the advanced response fine-tunes the weights of the fully connected layer of the model, and enables historical state snapshots when switching high-frequency parameters to prevent feature drift.

[0187] Step S7: Multi-level diagnosis is initiated when a machining anomaly occurs: Spot distortion accompanied by a jump in working distance is identified as a defocus fault; when ablation anomalies coexist with wavelength shift, material absorption spectrum conflicts are identified; and for systemic failures, a cross-entity traceability tree is constructed to locate the root cause of heat accumulation. Successful diagnoses increase the confidence level of the relationship edge, while repeated misdiagnoses trigger retraining.

[0188] Step S8: The final material verification phase compares the micro-area spectral characteristics: After identifying unknown impurities, the metal type switches to the high-energy pulse blasting template, and the organic pollutants activate the low-temperature long pulse cleaning template.

[0189] Step S2 also includes:

[0190] High-fluctuation state marking mechanism: When the pulse width fluctuation variance exceeds the threshold three times in a row, the time domain entity collaborative constraint reinforcement is activated;

[0191] Spot distortion weight correction: Dynamically adjust the weight coefficient based on the product of the working distance jump amplitude and the spot ellipticity increment;

[0192] Material mapping library reconstruction: retain the basic topology in the time and space domains, and only replace the frequency domain absorption spectrum eigenvectors.

[0193] Step S3 also includes:

[0194] Conditional activation of the spatial transformation layer: the spot ellipticity is used as the only basis for start and stop judgment, avoiding full-time calculation;

[0195] Cross-attention dynamic focusing: when the pulse frequency is greater than the critical value and the spot diameter is less than the characteristic size, the energy density area weight is automatically assigned;

[0196] Reflectance Compensation Scaling: The scaling factor is proportional to the wavelength reflectance confidence score.

[0197] Step S4 also includes:

[0198] Rule-based hard interception: Establish a wavelength material conflict rule library, and polymer materials trigger absolute interception in the near-infrared band;

[0199] Perturbation channel management: Femtosecond pulse activation perturbation amplitude compression algorithm, brittle material scenario locking repetition frequency perturbation dimension.

[0200] Step S5 also includes:

[0201] Same-direction deviation compensation: three consecutive cycles of deviation trigger a step-by-step compensation of the repetition frequency;

[0202] Spectral verification mechanism: When the measured surface roughness value exceeds the tolerance band predicted by the model, laser-induced breakdown spectroscopy detection is initiated.

[0203] Step S6 also includes:

[0204] Incremental update priority: When removing the threshold of ablation depth error metamaterial single pulse, the atlas coupling coefficient is forced to be updated first;

[0205] Historical snapshot anchoring: Save the mean value of the pulse cluster features before the parameter switching as the constraint benchmark for the weight update direction.

[0206] Step S7 also includes:

[0207] Tracing tree construction algorithm: Taking the abnormal node as the root, the diagnostic path is generated in descending order of the confidence of the relationship edge;

[0208] Dynamic confidence rewards and penalties: Successful diagnosis receives an additional score equal to the inverse of the time required to resolve the fault, while misdiagnosis decays exponentially based on the number of consecutive times.

[0209] The sensor acquisition in step S1 adopts conventional photoelectric conversion technology, and the emergency template in step S8 calls the standard parameter preset library. All threshold settings comply with international laser processing safety standards.

[0210] Through the deep collaboration of dynamic knowledge graphs and deep learning models, the precise analysis and real-time optimization of the coupling relationship between multi-domain parameters of pulsed lasers are achieved. Through the cross-domain coupling quantification technology of the dynamic knowledge graph, the interactive effects of time, space, and frequency domain parameters are converted into a coupling coefficient matrix that can be updated online, solving the problem that traditional methods have difficulty in dynamically analyzing strong parameter coupling. Through the working condition-triggered computing architecture of the hybrid model, submodules are dynamically activated based on the LSTM-CNN-GNN heterogeneous network, significantly reducing the computational load while ensuring prediction accuracy. Based on a dual closed-loop verification and incremental learning system, physical conflict parameters are intercepted through knowledge graph rule verification and adversarial robustness testing, and combined with an error-driven hierarchical update strategy to ensure continuous system optimization.

[0211] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0212] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A pulsed laser control system integrating dynamic knowledge graph and deep learning, characterized by: Including sequential connections: Dynamic knowledge graph construction module, deep learning hybrid model module and dual closed-loop collaborative optimization module; The input end of the dynamic knowledge graph construction module is connected to the time domain sensor, spatial domain sensor and frequency domain sensor of the pulse laser system, which is used to collect pulse width, repetition frequency, spot size, working distance and laser wavelength parameters in real time; The input end of the deep learning hybrid model module is connected to the output end of the dynamic knowledge graph construction module, which is used to process multi-domain parameters and output predicted values ​​of machining accuracy, ablation depth and surface quality; The input end of the dual closed-loop collaborative optimization module is connected to the output end of the deep learning hybrid model module, the control end thereof is connected to the actuator of the pulse laser system, and the feedback end thereof is connected to the output end of the sensor.

2. The pulse laser control system integrating dynamic knowledge graph and deep learning according to claim 1, characterized in that: The dynamic knowledge graph construction module includes: An entity relationship definition unit associates time domain parameters with pulse width and repetition frequency entities, spatial domain parameters with spot size and working distance entities, and frequency domain parameters with laser wavelength entities; Cross-domain coupling strength quantification unit generates dynamic coupling coefficient matrices between time domain and space domain, time domain and frequency domain, and space domain and frequency domain; The real-time update unit dynamically adjusts the entity node relationship and coupling coefficient matrix according to the sensor data.

3. The pulse laser control system integrating dynamic knowledge graph and deep learning according to claim 2, characterized in that: The cross-domain coupling strength quantization unit operates in the following manner: Initialize the coupling coefficient matrix based on the pulse laser transmission equation; The coupling coefficient matrix is ​​updated through training of historical data collected by the photodetector and the spot analyzer; When abnormal machining accuracy is detected, real-time correction of the coupling coefficient matrix is ​​triggered.

4. The pulse laser control system integrating dynamic knowledge graph and deep learning according to claim 1, characterized in that: The deep learning hybrid model module includes parallel connections: The time series processing unit uses a long short-term memory network with a temporal convolution kernel to process the sequence characteristics of pulse width and repetition frequency; The spatial processing unit uses a convolutional neural network with a spatial transformation layer to process the spatial distribution characteristics of the spot size and working distance; The frequency domain processing unit uses a graph neural network to embed the feature vectors of laser wavelength nodes in the knowledge graph.

5. The pulse laser control system integrating dynamic knowledge graph and deep learning according to claim 4, characterized in that: The output end of the deep learning hybrid model module is connected to a dual verification unit, which includes: Knowledge graph rule validator, which verifies whether the parameter combination violates the entity relationship constraints; Adversarial robustness tester that generates perturbations in the neighborhood of optimized parameters and detects performance fluctuations.

6. The pulse laser control system integrating dynamic knowledge graph and deep learning according to claim 5, characterized in that: The dual verification unit performs: If a parameter combination is judged invalid by the knowledge graph rule validator, the combination is discarded; If the adversarial robustness tester detects that the performance fluctuation exceeds the threshold, it triggers model weight fine-tuning.

7. The pulse laser control system integrating dynamic knowledge graph and deep learning according to claim 1, characterized in that: The dual closed-loop collaborative optimization module includes: Parameter optimizer, which uses evolutionary algorithms to search for the optimal parameter combination that meets the target performance; An execution controller outputs the verified parameter combination to an actuator of the pulse laser system; The incremental learner updates the knowledge graph coupling coefficient matrix and model weights based on the real-time performance data fed back by the sensors.

8. The pulse laser control system integrating dynamic knowledge graph and deep learning according to claim 7, characterized in that: The incremental learner operates as follows: When the deviation between the actual machining accuracy and the model prediction value exceeds the limit, the coupling coefficient matrix of the knowledge graph is updated first; If the deviation persists, an incremental learning algorithm is used to adjust the weights of the fully connected layers of the deep learning hybrid model.

9. The pulse laser control system integrating dynamic knowledge graph and deep learning according to claim 1, characterized in that: Also includes: The fault diagnosis unit, whose input is connected to the dynamic knowledge graph construction module, locates the cross-domain coupling fault source by backtracing the entity relationship path when the processing accuracy is abnormal.

10. A pulse laser control method based on any system of claims 1-9, characterized in that: The steps include: Step a: Collect time domain, spatial domain, and frequency domain parameters in real time through sensors; Step b: Dynamically update the entity relationships and coupling coefficient matrix of the knowledge graph; Step c: predict the system performance of the parameter combination through deep learning hybrid model; Step d: Outputting the optimal parameters to the pulse laser system after double verification; Step e: Update the knowledge graph and model weights based on real-time performance feedback.

Images