A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine
By using an intelligent closed-loop compensation system to monitor and analyze the thermal drift of fiber optic lenses in real time, the cutting quality problem caused by lens thermal drift in high-power fiber laser cutting machines is solved, achieving high-precision automatic compensation and stable control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU DAYE PHOTOELECTRIC CO LTD
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-30
Smart Images
Figure CN122299218A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of high-power fiber laser cutting machines, and more particularly to a closed-loop compensation system for thermal drift of fiber optic lenses in cutting machines. Background Technology
[0002] In the precision manufacturing field of high-power fiber laser cutting, the thermal drift problem of fiber lenses (especially focusing lenses) has become a core bottleneck restricting the performance and processing accuracy of high-end cutting equipment. This phenomenon stems directly from the physical characteristics of the core components under extreme operating conditions: when a high-energy beam of several kilowatts or even tens of thousands of kilowatts output by the laser continuously passes through the optical path system, the cumulative thermal effect caused by the absorption of trace amounts of radiation energy by the optical lenses cannot be completely avoided. This leads to microscopic thermal expansion and refractive index changes in the lenses, which in turn causes a slow, continuous, and nonlinear shift in the laser focal position. In actual operation, this microscopic physical change is directly amplified into macroscopic processing quality defects. Specifically, as the continuous processing time increases, the kerf width of the cutting surface gradually becomes uncontrollable, the perpendicularity deteriorates, the surface roughness increases significantly, and even intermittent cutting or slag buildup occurs.
[0003] Currently, industry solutions addressing this product-level pain point mainly fall into three categories, but all have inherent limitations. First, there's the common passive thermal management combined with open-loop compensation strategies, relying on enhanced cooling and preset parameter adjustments, which cannot respond in real-time to dynamic thermal disturbances during processing. Second, there are attempts to directly monitor lens temperature or deformation; however, in the complex and demanding optical path environment inside high-power lasers, integrating high-precision sensors is impractical and costly, and the mapping relationship between monitoring data and the final cutting morphology is difficult to accurately model. Third, there are methods that rely on monitoring indirect signals such as sparks and acoustic emissions for judgment; however, these signals have high noise levels and weak correlation with cross-sectional geometric quality, making precise closed-loop control impossible.
[0004] To address the aforementioned technical deficiencies, a closed-loop compensation system for thermal drift of fiber optic lenses in a fiber optic cutting machine is proposed. Summary of the Invention
[0005] The purpose of this invention is to enhance the overall intelligence of the device's autonomous perception and autonomous control of fiber optic lens decision-making compensation through its intelligent and closed-loop design.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: A closed-loop compensation system for thermal drift of fiber optic lenses in a fiber optic cutting machine includes: The dynamic acquisition unit collects dimensional maps of the product cross-section in real time within a preset period and sends them to the data storage unit in combination with the three-dimensional coordinate system and timestamp to form a spatiotemporal morphological feature map. The dynamic analysis unit acquires the spatiotemporal topographic feature maps of adjacent periods stored in the data storage unit and analyzes, compares and judges the spatiotemporal topographic feature results by combining them with the preset customized spatiotemporal topographic feature maps, and sends the spatiotemporal topographic feature results to the compensation processing unit. The compensation processing unit receives the spatiotemporal topographic feature results and matches and filters them with the preset thermal drift compensation parameters to generate a decision instruction package, and then sends the decision instruction to the compensation execution unit. The compensation execution unit receives dynamic hierarchical compensation instructions, performs control operations, and conducts re-inspection to realize the function of intelligent closed-loop re-judgment compensation.
[0007] Furthermore, the spatiotemporal morphological feature map is a spatiotemporal serialized morphological data structure constructed within a preset spatial region, based on a unified three-dimensional coordinate system and combined with timestamps.
[0008] Furthermore, the spatiotemporal topographic feature map includes at least: A1. Spatial index information used to characterize a preset spatial region ROI in a unified three-dimensional coordinate system; A2. Morphological feature data used to characterize the cross-sectional morphology corresponding to each sampling period; A3. Timestamp information used to characterize the time of data collection for morphological features.
[0009] Furthermore, the specific steps for generating the spatiotemporal topographic feature results are as follows: Using a pre-defined, customized spatiotemporal topography feature map as a benchmark, coordinate unification processing is performed on the spatiotemporal topography feature maps within adjacent periods, and adjacent standardized spatiotemporal topography feature maps are generated respectively. The dimensional deviation calculation is performed on the two adjacent standardized spatiotemporal topography feature maps to obtain the adjacent deviation value. Then, the dimensional deviation calculation is performed on the adjacent standardized spatiotemporal topography feature map to obtain the first benchmark deviation and the second benchmark deviation. The adjacent deviation value, the first benchmark deviation, and the second benchmark deviation are normalized to obtain the average deviation value. When the average deviation value is less than or equal to the preset deviation value, the equipment operates normally. When the average deviation value is greater than the preset deviation value, the two adjacent standardized spatiotemporal topography feature maps and the pre-defined, customized spatiotemporal topography feature map are input into the multimodal computing model for separate processing and matching of topography feature decision information. The matched topography feature decision information is marked as the spatiotemporal topography feature result.
[0010] Furthermore, the specific calculation process for dimensional deviation is as follows: Within a predefined spatial region, there is a uniform two-dimensional planar coordinate grid with m×n grid nodes. The depth value of any node (i,j) in the three-dimensional coordinate system is denoted as Zij. For any two spatiotemporal topographic feature maps GA and GB, the dimensional deviation value D(A,B) is defined as the root mean square error of the difference between the height values of all corresponding nodes in the two maps, and the calculation formula is as follows: , Where Zij(A) and Zij(B) represent the depth values of feature maps GA and GB at node (i,j), respectively.
[0011] Furthermore, the multimodal calculation model includes a cross-sectional profile calculation model, a kerf width layer calculation model, a taper layer calculation model, a verticality layer calculation model, a stripe layer calculation model, a rough layer calculation model, a slag layer calculation model, and an asymmetric layer calculation model.
[0012] Furthermore, the spatiotemporal morphological characteristics include the specific defect type, deviation value, location of occurrence, and confidence level of the cut section.
[0013] Furthermore, the specific generation process of the spatiotemporal topographic feature results: Two adjacent standardized spatiotemporal topographic feature maps and a pre-defined customized spatiotemporal topographic feature map are input into the sub-models of the initialized multimodal computing model for processing. The input feature maps are quantitatively analyzed from their own dimensions, and the corresponding topographic feature values and vectors are output. Then, by integrating the morphological feature data output from all multimodal computational models, a comprehensive multidimensional morphological feature vector is generated for the feature map analyzed for each sub-model. The difference between this vector and the baseline is then generated. The system accurately locates the defect type, quantifies the deviation, assesses the confidence level, and records its spatiotemporal location, thereby generating a structured spatiotemporal morphological feature result set R={r1,r2,...,rn}, where each sub-model result includes a defect type identifier, deviation quantification value, confidence level, and location coordinates.
[0014] Furthermore, the specific process of generating the decision instruction package through matching and filtering is as follows: First, based on the spatiotemporal topographic feature result set R={r1,r2,...,rn}, a knowledge base of pre-defined rules for matching topographic defects with thermal drift root causes is matched. Then, the matching degree is calculated to filter out one or more thermal drift root causes. For multiple thermal drift root causes, all feature evidence is comprehensively evaluated, and probability calculation and priority ranking are performed to determine the dominant root cause. Once the dominant root cause is determined, the corresponding compensation parameter model is called to convert the topographic deviation value into precise equipment adjustment parameters. Finally, all equipment adjustment parameters are encapsulated into a structured, executable decision instruction package.
[0015] Furthermore, the decision instruction package includes diagnostic conclusions, a list of compensatory actions, expected effects, and safety verification results.
[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This invention constructs an intelligent closed loop encompassing "collection, analysis, decision-making, execution, and re-inspection." Firstly, based on spatiotemporal morphology feature maps, it transforms traditional, static cross-section detection into dynamic morphology monitoring that integrates time series within a unified three-dimensional coordinate system. This not only captures instantaneous defects but also quantifies the morphology drift trend due to thermal accumulation, providing a data foundation for early warning and intervention. Secondly, it employs a two-level diagnostic strategy of "rapid coarse judgment + multimodal fine analysis." Through efficient dimensional deviation calculations, it performs real-time health screening, initiating refined multimodal model analysis only in cases of anomalies. This ensures diagnostic depth, accurately locating specific defect types such as kerf width, taper, and roughness, as well as quantifying deviations, while significantly reducing the system's computational load and response latency. Finally, the system achieves a fully automated decision-making chain from "identifying morphological anomalies" to "locking down the root cause of thermal drift" and then "outputting precise compensation commands." This transforms the previously experience-dependent, lagging process adjustments into a forward-looking closed-loop compensation based on real-time data, thereby significantly improving the dimensional consistency, morphological regularity, and process stability of the cut cross-section.
[0017] This invention ensures the effectiveness of compensation actions through a closed-loop re-judgment compensation mechanism. The system verifies the effect by re-checking the cross-section of the next cycle after execution, and decides whether to maintain, fine-tune, or re-determine based on the results, forming a continuously optimizing adaptive control loop. This significantly reduces the risk of scrap due to improper or over-compensation. This invention possesses strong self-learning and expansion potential through its knowledge-based driven mode. The preset thermal drift compensation parameter knowledge base can be customized and accumulated according to different materials and thicknesses, making the compensation strategy increasingly accurate with use. Simultaneously, the modular multimodal calculation model allows users to flexibly edit, add, or delete analysis dimensions according to actual needs, enabling the system to quickly adapt to new processing requirements or quality standards.
[0018] In summary, this invention directly improves the processing accuracy and product qualification rate of laser cutting machines during long-term operation, reduces downtime for adjustments and material waste, and further enhances the intelligence of autonomous perception and autonomous control of fiber optic lens decision compensation through its intelligent and closed-loop design. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings. Figure 1 This is a schematic diagram of the unit information flow of the present invention; Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] Example: A closed-loop compensation system for thermal drift of fiber optic lenses in a fiber optic cutting machine includes: The dynamic acquisition unit collects dimensional maps of the product cross-section in real time within a preset period and sends them to the data storage unit in combination with the three-dimensional coordinate system and timestamp to form a spatiotemporal morphological feature map. Product cross-section dimension map: used to characterize the geometric / morphological information of the cut cross-section area, which can be obtained by side-view imaging, line laser profile, cross-section surrogate quantity of coaxial vision or other sensing methods; the dimension map may include one or more of the following dimensions: cross-section profile, kerf width, taper, perpendicularity, stripes, roughness, slag, asymmetry, etc.
[0022] Spatiotemporal topographic feature map is a spatiotemporal serialized topographic data structure constructed within a preset spatial region, based on a unified three-dimensional coordinate system and combined with timestamps. The spatiotemporal topographic feature map includes at least: A1. Spatial index information used to characterize a preset spatial region ROI in a unified three-dimensional coordinate system; A2. Morphological feature data used to characterize the cross-sectional morphology corresponding to each sampling period; A3. Timestamp information used to characterize the time of data collection for morphological features.
[0023] Within a predefined spatial region, a spatiotemporal serialized topographic data structure is constructed based on a unified three-dimensional coordinate system and combined with timestamps, including at least ROI spatial index information, topographic dimension data corresponding to the sampling period, and timestamp information.
[0024] The dynamic analysis unit acquires the spatiotemporal topographic feature maps of adjacent periods stored in the data storage unit and analyzes and compares them with the preset customized spatiotemporal topographic feature maps to generate spatiotemporal topographic feature results, and sends the spatiotemporal topographic feature results to the compensation processing unit. The preset customized spatiotemporal topographic feature map is a reference template or benchmark space map under a specific material / thickness / process window. It is used to constrain the coordinate system, spatial scale and the boundary of the preset spatial region, and to calculate the template deviation to determine whether the current cross-sectional quality deviates from the target range.
[0025] The specific steps for generating spatiotemporal topographic feature results are as follows: Using a pre-defined, customized spatiotemporal topography feature map as a benchmark, coordinate unification processing is performed on the spatiotemporal topography feature maps within adjacent periods, generating adjacent standardized spatiotemporal topography feature maps. Dimensional deviation calculations are performed on the two adjacent standardized spatiotemporal topography feature maps to obtain adjacent deviation values. These are then compared with the pre-defined, customized spatiotemporal topography feature map to obtain a first reference deviation and a second reference deviation. The adjacent deviation values, the first reference deviation, and the second reference deviation are normalized to obtain an average deviation value. When the average deviation value is less than or equal to the pre-defined deviation value, the equipment operates normally. When the average deviation value is greater than the pre-defined deviation value, the two adjacent standardized spatiotemporal topography feature maps and the pre-defined, customized spatiotemporal topography feature map are input into a multimodal computing model for separate processing and matching of topography feature decision information. The matched topography feature decision information is marked as the spatiotemporal topography feature result. The coarse deviation threshold judgment uses a uniform dimensional deviation value to quickly determine "whether adjustment is needed." Only when the average deviation value exceeds the limit is the multimodal fine judgment model activated to output a specific problem type, reducing false triggering and computational burden.
[0026] The specific calculation process for dimensional deviation is as follows: Within a predefined spatial region, there is a uniform two-dimensional planar coordinate grid with m×n grid nodes. The depth value of any node (i,j) in the three-dimensional coordinate system is denoted as Zij. For any two spatiotemporal topographic feature maps GA and GB, the dimensional deviation value D(A,B) is defined as the root mean square error of the difference between the height values of all corresponding nodes in the two maps, and the calculation formula is as follows:
[0027] Where Zij(A) and Zij(B) represent the depth values of feature maps GA and GB at node (i,j), respectively.
[0028] In actual calculations, adjacent standardized spatiotemporal topographic feature maps and preset customized spatiotemporal topographic feature maps are marked and substituted into the above root mean square error formula to calculate the adjacent deviation value, the first reference deviation and the second reference deviation respectively.
[0029] The specific generation process of spatiotemporal topographic feature results: Two adjacent standardized spatiotemporal topographic feature maps and a pre-defined customized spatiotemporal topographic feature map are input into the sub-models of the initialized multimodal computing model for processing. The input feature maps are quantitatively analyzed from their own dimensions, and the corresponding topographic feature values and vectors are output. Then, by integrating the morphological feature data output from all multimodal computational models, a comprehensive multidimensional morphological feature vector is generated for the feature map analyzed for each sub-model. The difference between this vector and the baseline is then generated. The system accurately locates the defect type, quantifies the deviation, assesses the confidence level, and records its spatiotemporal location, thereby generating a structured spatiotemporal morphological feature result set R={r1,r2,...,rn}. Each sub-model result includes the defect type identifier, deviation quantification value, confidence level, location coordinates, etc. That is, a single spatiotemporal morphological feature result includes the specific defect type, deviation value, occurrence location, confidence level, etc. of the cutting section. The multimodal computational models include cross-sectional contour calculation models, kerf width layer calculation models, taper layer calculation models, perpendicularity layer calculation models, stripe layer calculation models, roughness layer calculation models, slag layer calculation models, and asymmetry layer calculation models. For example, cross-sectional contour calculation can use B-spline curve fitting or the RANSAC algorithm; kerf width calculation can be based on edge detection and distance transformation; taper and perpendicularity analysis can use linear regression and principal component analysis methods, respectively; stripe feature analysis can be performed using Fourier transform or gray-level co-occurrence matrix; roughness evaluation can directly use three-dimensional surface roughness parameters; slag recognition can combine image segmentation and morphological processing; asymmetry quantification can be based on contour moment analysis, and traditional edge detection can be upgraded to a semantic segmentation model based on deep learning to improve the recognition accuracy of complex slag.
[0030] These algorithms are all well-known and mature technologies in the fields of computer vision, digital image processing, and industrial metrology. Furthermore, these models can be modified, edited, and expanded according to actual needs, processing materials, and quality indicators. For specific requirements regarding processing materials such as steel plates, glass, and composite materials, and quality indicators such as heat-affected zone width and molten pool morphology, existing model algorithms can be replaced, parameters adjusted, or entirely new dedicated analysis modules can be added.
[0031] The compensation processing unit receives the spatiotemporal topographic feature results and matches and filters them with the preset thermal drift compensation parameters to generate a decision instruction package, and then sends the decision instruction to the compensation execution unit. The specific process of generating decision instruction packages through matching and filtering is as follows: First, a knowledge base of pre-defined rules for matching morphological defects with thermal drift root causes is established based on the spatiotemporal morphological feature results. Then, the matching degree is calculated to filter out one or more thermal drift root causes. For multiple thermal drift root causes, all feature evidence is comprehensively evaluated, and probability calculation and priority ranking are performed to determine the dominant root cause. Once the dominant root cause is determined, the corresponding compensation parameter model is called to convert the morphological deviation value into precise equipment adjustment parameters. Finally, all equipment adjustment parameters are encapsulated into a structured, executable decision instruction package, which includes diagnostic conclusions, a list of compensation actions, expected effects, and safety verification results.
[0032] The specific algorithm process for matching and filtering decision instruction packets is as follows: First, the compensation processing unit receives the spatiotemporal topographic feature result set R from the dynamic analysis unit. A pre-stored thermal drift compensation parameter knowledge base KB = {rule1, rule2, ..., rulem} defines the mapping relationship between defect features and root causes and compensation actions, as well as the rule weights, priorities, and compensation functions for each rule j = (Conditionj, Causej, Actionj, Weightj, Priorityj, Φj).
[0033] Secondly, feature parsing and root cause mapping are performed. For each feature result ri∈ R, Extract its key feature vector Fi = Encode(Ti, Sign(Vi), Quantize(Li)).
[0034] Then, iterate through each rule j in the knowledge base KB and calculate the matching degree:
[0035] For all matching rules, classify them according to their root cause Causej, and cumulatively calculate the activation intensity A(causek) of each root cause causek.
[0036] Next, multi-parameter collaborative screening and conflict resolution are performed: Synergistic effects among multiple feature results are used to update the confidence score of each root cause:
[0037] Then, the probability distribution of each root cause is calculated using the Softmax function: , Finally, the dominant root cause was selected:
[0038] The process then moves to the parameterization stage of the compensation decision. Here, γ is a preset conflict penalty weight coefficient, used to balance diagnostic confidence with the coordination of compensation actions when finally determining the dominant compensation root cause, ensuring that the generated decision instruction package is safe, conflict-free, and can be reliably implemented.
[0039] Based on the identified dominant root cause, the corresponding compensation function Φd and the current equipment state S are invoked to calculate the original compensation amount Δ=Φd(V,S,Θ), where V is a vector composed of all deviation values. Subsequently, a safety boundary constraint is applied to the compensation amount: Δconstrained=Clip(Δ,Δmin,Δmax).
[0040] V is a vector consisting of the deviation quantification values of all identified morphological defects. It is the direct input factual basis for compensation calculation and is used to quantify the specific differences between the current cross-section and the ideal morphology in multiple dimensions.
[0041] S is a vector representing the real-time working state of the cutting machine system. It serves as the contextual input parameter for compensation decisions, ensuring that the compensation command matches the actual working conditions such as the current laser power, lens temperature, and actuator position, so that the decision fits the instantaneous state of the system.
[0042] Θ is a set of internal parameters (such as proportional coefficients and bias terms) of the compensation function stored in the knowledge base and bound to a specific root cause. As the core of the empirical model that converts deviations into actions, its accuracy directly determines the precision of the compensation calculation.
[0043] When Φd(V,S,Θ) represents the root cause diagnosed, a preliminary adjustment plan is calculated based on the specific morphological deviation (V) and real-time equipment status (S) using a calibrated compensation model (Θ).
[0044] Δ is the original compensation command vector calculated by the compensation function, which directly indicates the theoretically required adjustment amount for each actuator, such as the focal length motor and the power controller.
[0045] Δmin and Δmax are the minimum and maximum allowable variation boundaries preset by the system for each compensation quantity. They serve as hard constraints to ensure operational safety and are set based on the physical limits of the equipment and the process safety window.
[0046] By using the Clip(Δ,Δ_min,Δ_max) operation, the theoretically calculated original compensation amount Δ is strictly limited within the safety boundaries Δmin and Δmax, thereby generating the final constrained compensation amount that can be executed safely and reliably, ensuring that the compensation action is effective and will never damage the equipment or cause process risks.
[0047] Finally, the decision instruction package is assembled and sent. The decision instruction package includes a unique instruction ID, timestamp, diagnostic root cause, a list of associated feature result IDs, a list of specific compensation actions, a predicted expected effect, and a security verification result. When the decision instruction package is sent to the compensation execution unit, it drives the physical actuator to complete closed-loop compensation. The entire matching, filtering, and generation process completes intelligent decision-making from "feature diagnosis" to "precise execution."
[0048] The compensation execution unit receives dynamic hierarchical compensation instructions, performs control operations, and conducts re-inspection to realize the function of intelligent closed-loop re-judgment compensation.
[0049] After the control operation waits for the processing to complete a full preset cycle, it repeats the above steps and determines whether fine-tuning iterative compensation is needed. If the deviation does not improve or worsens after at least n judgments, the compensation strategy is deemed to have failed, triggering an anomaly diagnosis process or manual intervention. Finally, regardless of the outcome, the complete closed-loop data of "execution-collection-reassessment-decision" is recorded for effect tracking and knowledge base optimization, thereby achieving continuous verification and self-optimization of the thermal drift compensation effect. n is a settable positive integer, typically 3-5 times.
[0050] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.
[0051] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine, characterized in that, include: The dynamic acquisition unit collects dimensional maps of the product cross-section in real time within a preset period and sends them to the data storage unit in combination with the three-dimensional coordinate system and timestamp to form a spatiotemporal morphological feature map. The dynamic analysis unit acquires the spatiotemporal topographic feature maps of adjacent periods stored in the data storage unit and analyzes, compares and judges the spatiotemporal topographic feature results in combination with the preset customized spatiotemporal topographic feature maps, and sends the spatiotemporal topographic feature results to the compensation processing unit. The compensation processing unit receives the spatiotemporal topographic feature results and matches and filters them with the preset thermal drift compensation parameters to generate a decision instruction package, and then sends the decision instruction to the compensation execution unit. The compensation execution unit receives dynamic hierarchical compensation instructions, performs control operations, and conducts re-examination to realize the function of re-judgment compensation.
2. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine according to claim 1, characterized in that, The spatiotemporal topographic feature map is a spatiotemporal serialized topographic data structure constructed within a preset spatial region, based on a unified three-dimensional coordinate system and combined with timestamps.
3. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine according to claim 2, characterized in that, Spatiotemporal topographic feature maps should include at least: A1. Spatial index information used to characterize a preset spatial region ROI in a unified three-dimensional coordinate system; A2. Morphological feature data used to characterize the cross-sectional morphology corresponding to each sampling period; A3. Timestamp information used to characterize the time of data collection for morphological features.
4. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine according to claim 2, characterized in that, The specific steps for generating spatiotemporal topographic feature results are as follows: Using a pre-defined, customized spatiotemporal topography feature map as a benchmark, coordinate unification processing is performed on the spatiotemporal topography feature maps within adjacent periods, and adjacent standardized spatiotemporal topography feature maps are generated respectively. The dimensional deviation calculation is performed on the two adjacent standardized spatiotemporal topography feature maps to obtain the adjacent deviation value. Then, the dimensional deviation calculation is performed on the adjacent standardized spatiotemporal topography feature map to obtain the first benchmark deviation and the second benchmark deviation. The adjacent deviation value, the first benchmark deviation, and the second benchmark deviation are normalized to obtain the average deviation value. When the average deviation value is less than or equal to the preset deviation value, the equipment operates normally. When the average deviation value is greater than the preset deviation value, the two adjacent standardized spatiotemporal topography feature maps and the pre-defined, customized spatiotemporal topography feature map are input into the multimodal computing model for separate processing and matching of topography feature decision information. The matched topography feature decision information is marked as the spatiotemporal topography feature result.
5. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine according to claim 4, characterized in that, The specific calculation process for dimensional deviation is as follows: Within a predefined spatial region, there is an identical two-dimensional planar coordinate grid with m×n nodes. The depth value of any node (i,j) in the three-dimensional coordinate system is denoted as Z. ij ; For any two spatiotemporal topographic feature maps GA and GB, the dimensional deviation value D(A,B) is defined as the root mean square error of the difference between the height values of all corresponding nodes in the two maps, and the calculation formula is as follows: ， Z ij (A) Z ij (B) Let represent the depth values of feature maps GA and GB at node (i,j), respectively.
6. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine according to claim 4, characterized in that, The multimodal calculation model includes the cross-sectional profile calculation model, the kerf width layer calculation model, the taper layer calculation model, the verticality layer calculation model, the stripe layer calculation model, the roughness layer calculation model, the slag layer calculation model, and the asymmetric layer calculation model.
7. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine according to claim 6, characterized in that, The spatiotemporal topographic features include the specific defect type, deviation value, location of occurrence, and confidence level of the cut section.
8. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine according to claim 6, characterized in that, The specific generation process of spatiotemporal topographic feature results: Two adjacent standardized spatiotemporal topographic feature maps and a pre-defined customized spatiotemporal topographic feature map are input into the sub-models of the initialized multimodal computing model for processing. The input feature maps are quantitatively analyzed from their own dimensions, and the corresponding topographic feature values and vectors are output. Then, by integrating the morphological feature data output from all multimodal computational models, and for each sub-model's analyzed feature map, a comprehensive multi-dimensional morphological feature vector is generated to compare with the baseline. The system accurately locates the defect type, quantifies the deviation, assesses the confidence level, and records its spatiotemporal location, thereby generating a structured spatiotemporal morphological feature result set R={r1,r2,...,r n Each sub-model result includes a defect type identifier, deviation quantification value, confidence level, and location coordinates.
9. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine according to claim 8, characterized in that, The specific process of generating decision instruction packages through matching and filtering is as follows: First, based on the spatiotemporal topographic feature result set R={r1,r2,...,r... n The system matches a pre-defined knowledge base of rules corresponding to morphological defects and thermal drift root causes. It then calculates the matching degree to filter out one or more thermal drift root causes. For multiple thermal drift root causes, it comprehensively evaluates all feature evidence, performs probability calculations and priority ranking to determine the dominant root cause. Once the dominant root cause is determined, it calls the corresponding compensation parameter model to convert the morphological deviation value into precise equipment adjustment parameters. Finally, it encapsulates all equipment adjustment parameters into a structured, executable decision instruction package.
10. A closed-loop compensation system for thermal drift of fiber optic lenses in a cutting machine according to claim 7, characterized in that, The decision instruction package includes diagnostic conclusions, a list of compensatory actions, expected effects, and safety verification results.