Film hole detection system based on image recognition and detection method thereof

Abstract

The invention relates to a film hole detection system based on image recognition and a detection method thereof, and belongs to the technical field of intelligent manufacturing and laser processing process control, and the method comprises the following steps: S1, adopting a multi-mode sensing module to collect various physical signals of a processing area in real time so as to form a multi-dimensional data stream; s2, inputting the multi-dimensional data stream into a digital twinning and simulation module to solve total interface stress, and generating a three-dimensional transient stress field distribution diagram; s3, inputting the multi-dimensional data stream into a defect prediction and early warning module to generate a defect occurrence probability; s4, generating a graded early warning signal based on the defect occurrence probability and a preset graded early warning threshold value; s5, based on the total interface stress and the graded early warning signal, a pulse energy control law is adopted, and an adjusted pulse energy parameter is solved; and S6, the adjusted pulse energy parameters are input into a laser machining execution module to achieve closed-loop control over the laser machining process, and an unprecedented data basis is provided for deep analysis and accurate prediction.

Description

An image recognition-based air film pore detection system and method Technical Field

[0001] This invention relates to the field of intelligent manufacturing and laser processing control, specifically to an image recognition-based air film hole detection system and its detection method. Background Technology

[0002] In the field of advanced laser processing, traditional process control methods mainly rely on fixed parameter settings and post-processing inspection. This open-loop processing method has significant limitations when faced with complex physical processes and dynamically changing working conditions at the microscale, making it difficult to guarantee the stability and yield of high-quality processing.

[0003] The aforementioned situation and shortcomings mainly stem from the lag in sensing and control technologies; specifically, they are manifested in the following ways: the sensing dimension is singular, failing to fully characterize the processing state: traditional methods usually rely on a single sensor or offline observation, which cannot capture in real time and comprehensively the complex phenomena of multi-physical field coupling such as temperature field, stress field, and plasma generated when laser interacts with matter; this kind of data-driven, piecemeal approach limits the system's understanding of the processing state and fails to provide sufficient basis for precise control.

[0004] Lack of defect prediction capability and serious lag in quality control: Traditional quality control is a reactive measure, that is, after defects, such as air film pores, have formed, they are detected by means of metallographic analysis. This delayed detection cannot prevent the occurrence of defects, often leading to the scrapping of workpieces, resulting in waste of materials and time, and it is difficult to deal with fundamental process stability problems.

[0005] The control method is rigid and cannot dynamically adapt to changes in working conditions: fixed processing parameters based on experience cannot cope with the minute inhomogeneities inside the material or disturbances in the processing environment; when microscopic signs that tend to form defects appear during processing, traditional open-loop systems are unable to identify and respond, and cannot actively intervene to avoid risks, leading to loss of control of the processing process.

[0006] The end result is that quality control in laser processing has long relied on expert experience, making it difficult to form a forward-looking and adaptive scientific control system driven by data and physical models. This limits its application potential and production efficiency in the field of ultra-precision and high-quality manufacturing.

[0007] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to provide an image recognition-based air film pore detection system and method to solve the problems mentioned in the background art.

[0009] The technical solution of this invention includes the following steps: S1. Using a multimodal sensing module, various physical signals of the processing area are collected in real time to form a multidimensional data stream; S2. The multidimensional data stream is input into a digital twin and simulation module to calculate the total interface stress and generate a three-dimensional transient stress field distribution map; S3. The multidimensional data stream is input into a defect prediction and early warning module to generate the defect occurrence probability; S4. Based on the defect occurrence probability and a preset graded early warning threshold, a graded early warning signal is generated; S5. Based on the total interface stress and the graded early warning signal, a pulse energy control law is used to calculate the adjusted pulse energy parameters; S6. The adjusted pulse energy parameters are input into a laser processing execution module to achieve closed-loop control of the laser processing process.

[0010] Preferably, S1 specifically includes: acquiring temperature field information using a high-speed infrared thermal imager; acquiring micro-vibration information using a laser Doppler vibrometer; acquiring high-frequency acoustic emission information using an acoustic emission sensor array; acquiring plasma spectral information using a plasma spectrometer; and acquiring coating interface state information using an eddy current detection probe.

[0011] Preferably, S2 specifically includes: determining thermal shock stress based on temperature field information; inferring phase transition induced stress based on plasma spectral information; determining mechanical shock stress based on micro-vibration information; weighting and summing the thermal shock stress, phase transition induced stress, and mechanical shock stress to generate the total interface stress; the weighting coefficients used for the weighting and summing are initially obtained from molecular dynamics simulations and continuously optimized using an online learning algorithm.

[0012] Preferably, S4 specifically includes: generating a level 3 warning signal when the probability of a defect occurrence is greater than the third warning threshold and the coating interface state information undergoes a sudden change; generating a level 2 warning signal when the probability of a defect occurrence is greater than the second warning threshold but not greater than the third warning threshold and the plasma spectrum or temperature field information undergoes abnormal distortion; generating a level 1 warning signal when the probability of a defect occurrence is greater than the first warning threshold but not greater than the second warning threshold; and not generating a warning signal when the probability of a defect occurrence is not greater than the first warning threshold; wherein, the first warning threshold, the second warning threshold, and the third warning threshold are determined based on receiver operating characteristic curve analysis.

[0013] Preferably, S5 specifically includes: comparing the total interface stress with a preset material critical stress threshold to generate a normalized predicted total stress; combining the reference energy, geometric scaling factor and normalized predicted total stress, and adjusting them through the stress feedback gain coefficient to calculate the adjusted pulse energy parameter; the reference energy is calibrated according to the workpiece material and thickness; the geometric scaling factor and stress feedback gain coefficient are optimized and determined through a deep reinforcement learning algorithm.

[0014] Preferably, it further includes: in response to a secondary warning signal, instantaneously increasing the stress feedback gain coefficient to reduce energy input; and in response to a tertiary warning signal, setting the adjusted pulse energy parameter to zero to stop processing.

[0015] Preferably, the reward function of the deep reinforcement learning algorithm is designed to maximize material removal efficiency and penalize any predicted stress exceeding a safety threshold or by a graded warning signal.

[0016] An image recognition-based air film pore detection system includes the following modules: a multimodal sensing module for real-time acquisition of various physical signals from the processing area to form a multidimensional data stream; a digital twin and simulation module for receiving the multidimensional data stream, calculating the total interface stress, and generating a three-dimensional transient stress field distribution map; a defect prediction and early warning module for receiving the multidimensional data stream to generate the defect occurrence probability; an early warning generation module for generating graded early warning signals based on the defect occurrence probability and a preset graded early warning threshold; an adaptive control module for calculating the adjusted pulse energy parameters based on the total interface stress and the graded early warning signals using a pulse energy control law; and a laser processing execution module for receiving the adjusted pulse energy parameters to achieve closed-loop control of the laser processing process.

[0017] This invention provides an improved image recognition-based air film pore detection system and method, which has the following improvements and advantages compared with existing technologies: 1. The multimodal sensing module used in this solution simultaneously collects temperature field, micro-vibration, high-frequency acoustic emission, plasma spectrum, and coating interface state information to construct a multi-physics, high-dimensional data view of the laser-matter interaction region. This multi-dimensional data stream allows the system to transcend the limitations of any single sensor in understanding the processing state, providing an unprecedented data foundation for in-depth analysis and accurate prediction. 2. This solution creatively designs a parallel dual-path analysis architecture; one path is a digital twin and simulation module based on a physical model, and the other path is a data-driven defect prediction and early warning module. In this architecture, the total interface stress... The solution demonstrates profound physical insight, ensuring that the model has both a solid theoretical foundation and adaptability to actual working conditions, which is not available in existing static models; 3. The hierarchical early warning signal generation logic proposed in this scheme is a risk management strategy that is far more advanced than traditional threshold alarms; by combining the continuous defect occurrence probability output by the artificial intelligence model with the confirmation information from key physical sensors, accurate risk classification is achieved; this mechanism of probability prediction plus physical confirmation greatly improves the reliability of early warning, effectively avoids unnecessary downtime caused by model misjudgment or sensor noise, and improves production efficiency while ensuring safety; 4. The system can autonomously learn optimal control strategies that surpass human intuition through massive virtual experiments in a digital twin environment, so as to maximize material removal efficiency while ensuring processing quality. Attached Figure Description

[0018] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 is a flowchart of an image recognition-based air film pore detection method of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0020] Example 1: Referring to Figure 1, this invention provides an image recognition-based method for detecting air film pores, comprising the following steps: S1, using a multimodal sensing module to collect various physical signals of the processing area in real time to form a multidimensional data stream; S2, inputting the multidimensional data stream into a digital twin and simulation module to calculate the total interface stress and generate a three-dimensional transient stress field distribution map; S3, inputting the multidimensional data stream into a defect prediction and early warning module to generate a defect occurrence probability; S4, generating a graded early warning signal based on the defect occurrence probability and a preset graded early warning threshold; S5, using pulse energy... The control law is used to calculate the adjusted pulse energy parameters; S6, the adjusted pulse energy parameters are input into the laser processing execution module to achieve closed-loop control of the laser processing process; a method for detecting air film holes based on image recognition is proposed, which constructs a complete closed-loop control process; the method begins with step S1, in which a multimodal sensing module is deployed in the processing area; the purpose of the module is to capture, in real time and synchronously, various physical phenomena generated when laser interacts with matter; by collecting these signals, the system can obtain a comprehensive, high-dimensional snapshot of the processing state, forming a multi-dimensional data stream; expansion of the sensing dimension; this method The multimodal sensing module employed in this design simultaneously acquires information on temperature field, micro-vibration, high-frequency acoustic emission, plasma spectrum, and coating interface state, constructing a multi-physics, high-dimensional data view of the laser-matter interaction region. This multi-dimensional data stream allows the system to transcend the limitations of any single sensor in understanding the processing state, providing an unprecedented data foundation for in-depth analysis and accurate prediction. In step S2, the acquired multi-dimensional data stream is transmitted to a digital twin and simulation module. This module utilizes the input physical signals and, through a mechanical model based on the principle of physical superposition, performs real-time... The total interfacial stress at the coating-substrate interface at the processing point is calculated. The stress calculation result is not only a numerical value but is also used to generate a visualized three-dimensional transient stress field distribution map, providing an intuitive basis for process analysis. Simultaneously, in step S3, the multi-dimensional data stream is also fed into a defect prediction and early warning module. This module embeds a trained artificial intelligence model, whose purpose is to analyze complex patterns in the data stream to assess the likelihood of impending defects such as air film pores in real time and output a quantified defect occurrence probability value. As one implementation method, the artificial intelligence model can employ a long short-term memory network; the network input is at time... Data vector acquired by the multimodal sensing module and normalized It consists of:

[0021] in, At any time The data vector acquired by the multimodal sensing module and then normalized; : at any time; Peak temperature; The maximum temperature gradient obtained through thermal image recognition; : Morphological factors of the heat-affected zone obtained through thermal image recognition; Surface vibration velocity; Root mean square of acoustic emission signal; : Specific plasma spectral line intensity ratio; Eddy current detection signal value; each component represents the peak temperature, the maximum temperature gradient obtained from thermal image recognition, the morphological factor of the heat-affected zone obtained from thermal image recognition, the surface vibration velocity, the root mean square of the acoustic emission signal, the intensity ratio of a specific plasma spectral line, and the eddy current detection signal value, respectively; the LSTM network can be designed to contain two hidden layers, each with 128 neurons, followed by a layer with a loss rate of 0.5 to prevent overfitting, and then through a fully connected layer using the Sigmoid activation function, outputting a scalar value between 0 and 1, which is the probability of defect occurrence. The model can be pre-trained using an offline dataset containing thousands of processing passes. This dataset includes a complete multi-dimensional data stream for each pass and its corresponding processing quality label obtained through metallographic analysis, indicating the presence or absence of defects. In step S4, the defect occurrence probability generated in the previous step is used to drive an early warning system. The probability value is compared with a set of preset graded early warning thresholds. This step aims to transform continuous probability values ​​into discrete graded early warning signals with clear operational instructions, such as level one, level two, or level three alarms. In step S5, an adaptive control logic is executed. The core of this step is the application of a pulse energy control law. The purpose of the control law is to dynamically calculate a... The adjusted laser pulse energy parameters; this is a control loop combining feedforward and feedback, designed to actively intervene in the processing; in step S6, the control loop is closed; the adjusted pulse energy parameters calculated in step S5 are sent to the laser processing execution module; the module aims to precisely execute energy adjustment commands, thereby achieving real-time, dynamic closed-loop control of the laser processing process to suppress defect formation; this method achieves refined control of the laser processing process by constructing a complete closed loop from real-time sensing, dual-path parallel analysis, hierarchical decision-making to precise execution; its core technical effect is that it can predict and avoid risks before the macroscopic formation of defects by deeply analyzing microscopic physical signals, thereby significantly improving the stability of processing quality and yield, and reducing reliance on traditional hysteresis quality inspection.

[0022] S1 specifically includes: acquiring temperature field information using a high-speed infrared thermal imager; acquiring micro-vibration information using a laser Doppler vibrometer; acquiring high-frequency acoustic emission information using an acoustic emission sensor array; acquiring plasma spectral information using a plasma spectrometer; and acquiring coating interface state information using an eddy current detection probe. To further clarify the implementation of the multi-modal sensing module in step S1, the purpose of this embodiment is to capture multi-physics field information during laser processing by deploying a set of complementary sensors, providing a comprehensive and reliable data foundation for subsequent analysis and control. The high-speed infrared thermal imager is used to acquire temperature field information and identify and analyze the generated two-dimensional thermal image sequence. Its role is not only to monitor peak temperature, but more importantly, to identify and quantify the morphological characteristics of the processing point and its heat-affected zone in real time through image recognition algorithms, such as shape, symmetry, temperature gradient distribution, and the spatiotemporal evolution behavior of abnormal hot spots. For example, the algorithm is trained to identify asymmetric or unstable thermal field morphologies highly correlated with film pore defects. This image-based feature recognition can provide richer process information than a single temperature value, revealing the thermal shock effect of the material more profoundly. Laser Doppler... Vibration meters are used to collect micro-vibration information; their function is to measure the instantaneous vibration velocity of the material surface without contact, reflecting the intensity of the mechanical shock wave generated by material ablation. Acoustic emission sensor arrays are used to collect high-frequency acoustic emission information; their function is to detect the elastic waves released by the initiation and propagation of microcracks within the material, a key means of capturing early signs of damage. Plasma spectrometers are used to collect plasma spectral information; their function is to infer the phase transition process, elemental composition, and temperature of the material by analyzing the spectral characteristics of the plasma generated during processing, which is closely related to the volume change induced by the phase transition. Eddy current detection probes are used to collect coating interface state information; their function is to monitor the bonding state between the coating and the substrate in real time, effectively detecting the formation of macroscopic defects such as delamination or debonding. The gain technology of the sensor combination provides the system with a dataset that is far richer and more robust than information from a single sensor. The cross-validation capability of various physical signals greatly improves the accuracy and reliability of the system's understanding of the processing state, laying a solid data foundation for subsequent high-precision stress calculation and defect prediction, effectively avoiding control errors caused by misjudgments from single sensor information.

[0023] S2 specifically includes: determining thermal shock stress based on temperature field information; inferring phase transition-induced stress based on plasma spectral information; determining mechanical shock stress based on micro-vibration information; and performing a weighted summation of the thermal shock stress, phase transition-induced stress, and mechanical shock stress to generate the total interface stress. The weighting coefficients used in the weighted summation are initially obtained from molecular dynamics simulations and continuously optimized using an online learning algorithm. To further clarify the method for calculating the total interface stress in the digital twin and simulation module of step S2, the purpose of this embodiment is to establish a stress calculation model closely related to physical reality. The calculation of the total interface stress is based on the physical superposition principle. Among them, thermal shock stress... The determination is based on the temperature field information collected by a high-speed infrared thermal imager; This refers to the stress generated by the drastic thermal expansion and contraction of materials, and its function is to quantify the challenge that thermal effects pose to interfacial bonding; its value can be expressed by the formula:

[0024] in, The coefficient of thermal expansion of the material. For elastic modulus, The measured instantaneous peak temperature, The reference temperatures for the materials are all derived from a pre-set material property database or calibrated experimentally; phase transformation induced stress The inference is based on plasma spectral information collected by a plasma spectrometer; This refers to the stress caused by the volume change of a material during its transition from a solid to a liquid or plasma state. Its function is to assess the impact of the phase transition process on interfacial stability; stress is considered to be related to the rate of volume change. Proportional, that is The bulk modulus The material parameters are preset, and the volume change rate is indirectly inferred from spectral analysis data; the volume change rate can be obtained from the intensity of ion spectral lines of specific elements in the plasma spectrum. With atomic spectral line intensity The ratio is used to approximate the relationship, and this ratio is positively correlated with plasma temperature and ionization degree. To establish a quantitative relationship between phase transition induced stress and plasma spectrum, we use the following semi-empirical formula for calculation:

[0025] in, It is a dimensionless conversion coefficient obtained through experimental calibration, which implicitly represents the conversion relationship from spectral line ratio to volume change rate; mechanical impact stress The determination is based on the micro-vibration information collected by a laser Doppler vibration meter; This refers to the stress generated by the shock wave formed by the recoil pressure of material ablation. Its function is to measure the impact load of mechanical effects on the interface; the magnitude of the stress is related to the surface vibration velocity. and material acoustic impedance The product is directly proportional, that is ,in This is the acoustic impedance of the material; this parameter is a preset parameter for the material. Measured in real time by a laser Doppler vibration meter; to convert the proportional relationship into an equation, a dimensionless impact coupling coefficient is introduced. The coefficient characterizes the efficiency with which ablation recoil pressure is converted into mechanical shock waves within the material; therefore, the mechanical shock stress can be determined by the following formula:

[0026] in, The value of is typically between 0 and 1 and can be determined through initial impact calibration experiments. After each stress component is determined, the system will perform a weighted summation of the thermal shock stress, phase transformation induced stress, and mechanical shock stress to generate the final total interfacial stress. The purpose of weighted summation is to comprehensively evaluate the combined impact of multiple physical effects on interface stability; its mathematical model is as follows:

[0027] In the expression: Arbitrary processing position and at any time The total interfacial stress, in Pascals, is calculated using this formula. : These are the thermal shock, phase change induced, and mechanical shock stress components, respectively, with dimensions in Pascals (Pa), and their values ​​are derived from the calculations of the aforementioned components. : Dimensionless weighting coefficients, whose function is to adjust the contribution of each stress component in the total stress calculation; Arbitrary processing position; : at any time; : Thermal shock stress component; Phase transformation induced stress components; Mechanical impact stress components; the method of determining the weighting coefficients used for weighted summation reflects the adaptability of the model; The initial values ​​were obtained through molecular dynamics simulations, ensuring that the initial settings of the coefficients have a solid microscopic physical basis. During system operation, an online learning algorithm is employed to continuously optimize these coefficients by comparing model predictions with subsequent measured data. The online learning algorithm can employ a normalized least mean square algorithm. The system converts the coating interface state information collected by the eddy current detection probe into a measured reference value for stress feedback through a preset mapping function. At each time step, define the prediction error:

[0028] in, Prediction error; The measured reference value for stress feedback is derived from the eddy current detection signal. Total interfacial stress calculated by the model; weight vector Update according to the following rules:

[0029] in, The vector composed of each stress component is formed as follows: ; The square of the Euclidean norm of the stress component vector; The learning rate determines the convergence speed and stability; for example, it can be 0.1. It is a small positive number set to prevent the denominator from being zero; its dimensions are the same as... To remain consistent, the value is the square of the pressure, for example... By updating according to the above rules, the weighting coefficients can be adaptively adjusted based on the real-time prediction error. :exist The updated weight vector at each time step; :exist The weight vector at time step is formed as follows It provides a dynamic and adaptive stress calculation model; by decomposing the total stress into physical components strongly correlated with specific sensor data and introducing online optimizable weighting coefficients, the stress calculation not only has clear physical meaning but can also continuously learn and adjust during use. This allows it to more accurately reflect complex and variable real-world processing conditions than a fixed, static physical model, providing a more precise basis for subsequent control decisions. It also features a dual-path fusion mechanism for analysis and decision-making; this solution creatively designs a parallel dual-path analysis architecture; one path is a digital twin and simulation module based on the physical model, and the other is a data-driven defect prediction and early warning module; in this architecture, the calculation of the total interface stress reflects profound physical insights; its calculation is not a simple black-box model but a formula based on the principle of physical superposition.

[0030] in, Total interfacial stress Thermal shock stress Phase transformation induced stress, Mechanical impact stress Dimensionless weighting coefficient : Any processing position At any given time; the practical significance of the formula lies in transforming multiple physical measurements that are difficult to correlate directly into a unified mechanical quantity with a clear physical meaning, namely, the total interfacial stress. It characterizes the combined effect of various physical effects on the stability of material interfaces; among them, the weighting coefficient The method for determining the model is to obtain initial values ​​through molecular dynamics simulation and then optimize them using an online learning algorithm. This ensures that the model has both a solid theoretical foundation and the ability to adapt to actual working conditions, which is something that existing static models do not possess.

[0031] Example 2S4 specifically includes: generating a level 3 warning signal when the defect occurrence probability is greater than the third warning threshold and the coating interface state information undergoes a sudden change; generating a level 2 warning signal when the defect occurrence probability is greater than the second warning threshold but not greater than the third warning threshold and the plasma spectrum or temperature field information undergoes abnormal distortion; generating a level 1 warning signal when the defect occurrence probability is greater than the first warning threshold but not greater than the second warning threshold; and not generating a warning signal when the defect occurrence probability is not greater than the first warning threshold. The first, second, and third warning thresholds are determined based on receiver operation characteristic curve analysis. In this example, the hierarchical warning signal generation logic in step S4 aims to establish a multi-level risk response mechanism that is structurally clear, logically rigorous, and timely. The mechanism is entirely based on the defect occurrence probability output by the defect prediction module. This is combined with confirmation information from key sensors to trigger alarms of different levels; the rules of the graded early warning mechanism are preset as follows: when the probability of defect occurrence... Not greater than the first warning threshold When the system determines the current processing state is safe, it does not generate a warning signal; when the probability of a defect occurring... Greater than the first warning threshold And not greater than the second warning threshold When the system triggers a Level 1 warning signal, it is a suggestive warning, indicating only that the risk is trending upward; when the probability of defect occurrence... Greater than the second warning threshold And not greater than the third warning threshold In such cases, a high-probability state must be corroborated by other key physical sensors; the system will only confirm the authenticity of the risk and trigger a secondary warning signal if and only if the state is confirmed by identifiable anomalous distortions in the plasma spectrum or temperature field information; the anomalous distortion at the point is quantitatively defined as: the instantaneous value of the signal. Deviation from its time window Moving average within Exceed Multiple moving standard deviations That is, when the following conditions are met: When this occurs, it is determined to be an abnormal distortion; among which, the parameter For example, a value of 3 and a time window ; : Signals, such as the instantaneous values ​​of plasma spectral or temperature field information; : The moving average of the signal over a time window N; Time window, in units of data points or number of samples, for example, 100 data points, which can be preset according to the baseline noise level; : The moving standard deviation of the signal within a time window N; the level is interventional warning, which will initiate proactive process parameter adjustments; when the probability of defect occurrence... Greater than a critical third warning threshold At this time, the system considers there to be an extremely high risk of defects; the warning also requires final physical confirmation; the system will only trigger the highest level three warning signal when and only when the coating interface state information collected by the eddy current detection probe undergoes a sudden change, indicating that macroscopic separation has occurred or is about to occur, and the signal will cause the processing to stop immediately; "sudden change in coating interface state information" is quantitatively defined as: eddy current detection signal The time rate of change, the absolute value of the first difference, exceeds the preset mutation threshold. That is, when the following conditions are met: When a mutation occurs, it is considered to have occurred; threshold It can be calibrated experimentally under the critical state where coating debonding defects just occur; among which, : The current value of the eddy current detection signal; : The eddy current detection signal value at the previous moment; : The change in time, i.e., the time step; Preset mutation threshold; first warning threshold Second warning threshold Compared with the third warning threshold This refers to the probability of a defect occurring. Three key values ​​are used for segmentation; their function is to transform continuous probability predictions into discrete risk levels with clear operational significance; these thresholds are determined not by experience, but by receiver operating characteristic curve analysis of a large-scale verification dataset; the method aims to find a mathematically optimal balance between maximizing the defect detection rate and minimizing the false alarm rate, ensuring the scientific validity and reliability of the early warning system; the technical benefit of this embodiment is the establishment of a dual verification early warning mechanism combining probability prediction and physical confirmation, which combines sensitivity and reliability; it avoids the false alarms that may arise from relying solely on probability models and overcomes... This solution overcomes the problem of delayed response when relying solely on physical sensors. This tiered logic with confirmation mechanisms enables the system to respond appropriately to different levels of risk: alerting for low-risk situations, proactively intervening for medium-risk situations, and immediately halting operations for critical risks. This ensures processing safety while maximizing processing continuity and efficiency. The tiered early warning signal generation logic proposed in this solution is a far more advanced risk management strategy than traditional threshold alarms. It combines the probability of continuous defect occurrence output by the artificial intelligence model with confirmation information from key physical sensors, achieving precise risk tiering. For example, triggering a level-two warning requires not only the probability exceeding a threshold... Furthermore, abnormal distortions in the plasma spectrum or temperature field are required as corroborating evidence. This mechanism of probabilistic prediction combined with physical confirmation greatly improves the reliability of early warning, effectively avoids unnecessary shutdowns caused by model misjudgments or sensor noise, and improves production efficiency while ensuring safety.

[0032] Example 3S5 specifically includes: comparing the total interface stress with a preset material critical stress threshold to generate a normalized predicted total stress; combining the reference energy, geometric scaling factor, and normalized predicted total stress, and adjusting them through the stress feedback gain coefficient to calculate the adjusted pulse energy parameters; the reference energy is calibrated according to the workpiece material and thickness; the geometric scaling factor and stress feedback gain coefficient are optimized and determined through a deep reinforcement learning algorithm; in this example, the pulse energy control law in step S5 aims to define a mathematically clear and physically meaningful adaptive energy adjustment formula to achieve feedforward control of the processing process; to make the stress feedback universal, normalization processing is required, that is, the total interface stress calculated by the digital twin module... Compared with the preset material critical stress threshold Compare; This refers to the inherent maximum stress limit that a material can withstand, such as its yield strength. This parameter serves as a benchmark, with values ​​derived from material handbooks or calibrated experimentally. The comparison method is division to generate a normalized predicted total stress. The calculation formula is:

[0033] It is a dimensionless number that characterizes how close the currently predicted stress is to the material's failure limit. Total interfacial stress calculated by the digital twin module; : Preset material critical stress threshold, such as yield strength; system combined with reference energy A geometric scaling factor Compared with the above normalized predicted total stress And through a stress feedback gain coefficient Adjustments were made, and the adjusted pulse energy parameters were calculated. The mathematical expression for the control law is:

[0034] In the expression: After adjustment, it can be applied to any processing position. The pulse energy, in units of joules (J), is the final output of this control law; The reference energy, measured in joules (J), defines a base energy level without stress feedback. : A dimensionless geometric scaling factor that pre-distributes energy non-uniformly based on the geometry of the processing path; The dimensionless stress feedback gain coefficient determines the strength of the predicted stress's suppression of energy, i.e., the sensitivity of the feedback. Normalized total stress is predicted, dimensionless, and serves as the core feedback signal. Arbitrary processing position; key adjustable parameters in the control law are determined as follows: reference energy. It was calibrated through preliminary process experiments based on the material and thickness of the workpiece, ensuring the applicability of the basic energy; while the geometric scaling factor... The specific functional form and the optimal stress feedback gain coefficient The value of is determined by optimizing the deep reinforcement learning algorithm in a digital twin environment; this ensures the optimality of the control strategy under complex working conditions; by closely integrating the physical model with control theory, a forward-looking defect prevention mechanism is created; through normalization, the applicability of the control law is broadened; by introducing parameters optimized by deep reinforcement learning, the control strategy can reach an optimal balance point that is difficult for human experts to set; the control law can automatically and accurately reduce energy input the moment a high-stress risk area is identified, thereby eliminating its influence before the stress accumulates to a critical point sufficient to form a defect, significantly improving the intrinsic quality of the processing.

[0035] It also includes: in response to a secondary warning signal, instantaneously increasing the stress feedback gain coefficient to reduce energy input; in response to a tertiary warning signal, setting the adjusted pulse energy parameter to zero to stop processing; the reward function of the deep reinforcement learning algorithm is designed to maximize material removal efficiency and penalize any predicted stress exceeding a safety threshold or by the graded warning signal; in this embodiment, the control strategy in the adaptive control module is further refined and optimized; the purpose is to link the output of the warning system with the energy control law and clarify the optimization objective of deep reinforcement learning, thereby forming a more intelligent and safer integrated control system; firstly, based on the above control law, the embodiment adds direct response logic to the warning signal; in response to a secondary warning signal, the system instantaneously increases the stress feedback gain coefficient. ; The increase will make The term decreases sharply, resulting in the final pulse energy. The pulse energy parameter is significantly and rapidly reduced; this is a powerful but non-terminating intervention designed to quickly remove the system from the hazardous process window. In response to the Level 3 warning signal, the system takes the most decisive action, directly adjusting the pulse energy parameter. The forced setting to zero immediately stops processing, thus minimizing further damage to the workpiece when irreversible damage is detected; secondly, the embodiments clarify the optimization methods used. and The core design principle of the reward function in deep reinforcement learning algorithms is as follows: the reward function is designed as a multi-objective optimization function, aiming to maximize material removal efficiency, which represents the pursuit of processing economy; at the same time, to ensure processing quality and safety, the function will handle any predicted stress exceeding a safety threshold, such as... The situation may be penalized by tiered warning signals, with penalties imposed for any level 1, 2, or 3 alarm. This means that when the algorithm explores more efficient processing strategies, it will be negatively incentivized by triggering any warning or causing excessive simulated stress, thus guiding it to find a globally optimal processing strategy that balances efficiency, quality, and safety. As a preferred implementation, deep reinforcement learning algorithms can employ proximal policy optimization algorithms. The core elements are defined as follows: Agent: Control policy model, whose task is to output optimal control parameters; Environment: Composed of the digital twin and simulation modules of this invention; State space: At any given time... and arbitrary processing position status The multi-dimensional data stream and the currently calculated total interfacial stress and any processing position Common definition, that is ;in, At any time and arbitrary processing position The state; Multi-dimensional data flow; The total interfacial stress calculated so far; Arbitrary processing position; At any given time; Action space: the actions output by the agent. It is a vector containing two values, namely ,in For stress feedback gain coefficient, and It is used to define the geometry scaling factor. A key parameter, for example, when When it is a Gaussian function, It can be used as its standard deviation; where, geometric scaling factor Arbitrary processing position The function, during calculation, allows for arbitrary processing positions. It will be a feature length Normalization, that is, using As input, therefore its key parameters It is also a dimensionless quantity; The preset feature length, such as the total length of the processing path or the size of a specific region, is used to achieve normalization processing.

[0036] Through extensive iterative training in a digital twin environment, the PPO algorithm learns an optimal policy that adapts to the input state. Dynamically output the best action To maximize the cumulative reward defined by the reward function; combining the above features, the resulting synergistic gain effect constructs a two-layer control safety mechanism of regular optimization and emergency intervention; the deep reinforcement learning algorithm is responsible for continuously finding and executing an optimal energy control strategy that can reduce the occurrence of warnings from the source under normal processing conditions; while the forced intervention logic based on the warning signal serves as an independent, high-priority safety barrier, ensuring that even in extreme cases where the model prediction deviates or unexpected events occur, the system can still make the fastest and most correct safety response; this combination ensures that the system balances high performance and high safety; the forward-looking and adaptive nature of control execution; the core of this scheme lies in its pulse energy control law, given by the formula:

[0037] in, Adjust the pulse energy after adjustment. Reference energy, Geometric scaling factor Stress feedback gain coefficient Normalized total stress; the formula is physically consistent and realizable; its practical significance is that it establishes a negative feedback control loop based on the predicted stress; when the normalized stress... Approaching 1, meaning the predicted energy output is close to the material's critical damage threshold. It automatically and non-linearly, proactively reducing stress before it accumulates to a dangerous level, rather than remedying damage after it occurs. Crucially, core parameters in the control law, such as the geometric scaling factor, are also involved. With stress feedback gain coefficient The optimal control strategy is determined through deep reinforcement learning algorithms; this means that the system can learn the optimal control strategy beyond human intuition through massive virtual experiments in a digital twin environment, so as to maximize material removal efficiency while ensuring processing quality.

[0038] Example 4: An image recognition-based air film hole detection system includes the following modules: a multimodal sensing module for real-time acquisition of various physical signals from the processing area to form a multidimensional data stream; a digital twin and simulation module for receiving the multidimensional data stream, calculating the total interface stress, and generating a three-dimensional transient stress field distribution map; a defect prediction and early warning module for receiving the multidimensional data stream to generate the defect occurrence probability; an early warning generation module for generating a graded early warning signal based on the defect occurrence probability and a preset graded early warning threshold; an adaptive control module for calculating the adjusted pulse energy parameters using a pulse energy control law based on the total interface stress and the graded early warning signal; and a laser processing execution module for receiving the adjusted pulse energy parameters to achieve closed-loop control of the laser processing process. This example provides an implementation method for an image recognition-based air film hole detection system, which serves as the carrier for implementing the aforementioned method. The entire system is designed as a tightly coupled modular architecture, employing edge computing... The computing platform and FPGA hardware acceleration technology ensure that the entire closed-loop response time from signal acquisition to control command execution is less than 1 millisecond. The system includes: a multimodal sensing module, physically configured as a sensor array deployed in the processing area, which not only acquires various physical signals from the processing area in real time, but also has a built-in image recognition and processing unit for real-time analysis and feature extraction of images acquired by the infrared thermal imager, forming a multi-dimensional data stream including image recognition features; a digital twin and simulation module, which can be deployed on a high-performance edge computing unit, which receives the multi-dimensional data stream from the sensing module and calculates the total interface stress in real time based on the above physical model, while generating a three-dimensional transient stress field distribution map for display and recording; and a defect prediction and early warning module, the core of which is a pre-trained artificial intelligence model running on the edge computing unit, which receives the multi-dimensional data stream and analyzes its features to generate a real-time defect occurrence probability. A warning generation module, consisting of a set of logic rules embedded in the controller, is used to generate graded warning signals based on the defect occurrence probability output by the upstream module and preset graded warning thresholds, combined with the aforementioned confirmation mechanism. An adaptive control module is used to implement the system's core control algorithm. Its task is to calculate the adjusted pulse energy parameters based on the total interface stress calculated by the digital twin and simulation module and the graded warning signals issued by the warning generation module, using the pulse energy control law defined above. A laser processing execution module, including a laser, optical path system, and galvanometer, etc., is used to receive adjusted pulse energy parameters from an adaptive control module and precisely execute commands on the workpiece surface, thereby achieving closed-loop control of the laser processing process. The system transforms the traditional open-loop control laser processing process into a closed-loop control process with real-time sensing, analysis, decision-making, and adaptive capabilities, enabling higher levels of processing quality, consistency, and efficiency in industrial production. By constructing a complete and high-speed closed loop encompassing sensing, analysis, decision-making, and execution, laser processing is transformed from an experience-based craft into a precise science driven by data and physical models. This achieves a paradigm shift from defect detection to defect prediction and prevention, resulting in processing quality stability, yield, and the ability to handle complex processes that are difficult to achieve with existing technologies.

[0039] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting air film pores based on image recognition, characterized in that, The process includes the following steps: S1, using a multimodal sensing module to collect various physical signals from the processing area in real time to form a multidimensional data stream; S2, inputting the multidimensional data stream into a digital twin and simulation module to calculate the total interface stress and generate a three-dimensional transient stress field distribution map. S3. Input the multi-dimensional data stream into the defect prediction and early warning module to generate the probability of defect occurrence; S4. Generate graded early warning signals based on the probability of defect occurrence and the preset graded early warning threshold; S5. Based on the total interface stress and graded early warning signal, the pulse energy control law is used to calculate the adjusted pulse energy parameters; S6. The adjusted pulse energy parameters are input into the laser processing execution module to achieve closed-loop control of the laser processing process.

2. The method for detecting air film pores based on image recognition according to claim 1, characterized in that, S1 specifically includes: acquiring temperature field information using a high-speed infrared thermal imager; acquiring micro-vibration information using a laser Doppler vibration meter; acquiring high-frequency acoustic emission information using an acoustic emission sensor array; acquiring plasma spectral information using a plasma spectrometer; and acquiring coating interface state information using an eddy current detection probe.

3. The method for detecting air film pores based on image recognition according to claim 1, characterized in that, S2 specifically includes: determining thermal shock stress based on temperature field information; inferring phase transition induced stress based on plasma spectral information; determining mechanical shock stress based on micro-vibration information; and weighting and summing the thermal shock stress, phase transition induced stress, and mechanical shock stress to generate the total interface stress. The weighting coefficients used for the weighting and summing are initially obtained from molecular dynamics simulations and continuously optimized using an online learning algorithm.

4. The method for detecting air film pores based on image recognition according to claim 1, characterized in that, S4 specifically includes: generating a level 3 warning signal when the probability of a defect occurring is greater than the third warning threshold and the coating interface state information changes abruptly; generating a level 2 warning signal when the probability of a defect occurring is greater than the second warning threshold but not greater than the third warning threshold and the plasma spectrum or temperature field information undergoes abnormal distortion; generating a level 1 warning signal when the probability of a defect occurring is greater than the first warning threshold but not greater than the second warning threshold; and not generating a warning signal when the probability of a defect occurring is not greater than the first warning threshold. The first, second, and third warning thresholds are determined based on receiver operating characteristic curve analysis.

5. The method for detecting air film pores based on image recognition according to claim 1, characterized in that, S5 specifically includes: comparing the total interface stress with a preset material critical stress threshold to generate a normalized predicted total stress; combining the reference energy, geometric scaling factor, and normalized predicted total stress, and adjusting them through the stress feedback gain coefficient to calculate the adjusted pulse energy parameters; the reference energy is calibrated according to the workpiece material and thickness; the geometric scaling factor and stress feedback gain coefficient are optimized and determined through a deep reinforcement learning algorithm.

6. The method for detecting air film pores based on image recognition according to claim 5, characterized in that, Also includes: In response to the level 2 warning signal, the stress feedback gain coefficient is increased instantaneously to reduce energy input; In response to the level 3 warning signal, the adjusted pulse energy parameter is set to zero to stop processing.

7. The method for detecting air film pores based on image recognition according to claim 5, characterized in that, The reward function of the deep reinforcement learning algorithm is designed to maximize material removal efficiency and penalize any predicted stress exceeding a safety threshold or by a graded warning signal.

8. An image recognition-based air film pore detection system, based on the image recognition-based air film pore detection method according to any one of claims 1-7, characterized in that, It includes the following modules: a multimodal sensing module, used to acquire various physical signals from the processing area in real time to form a multidimensional data stream; The digital twin and simulation module is used to receive multi-dimensional data streams, calculate the total interface stress, and generate a three-dimensional transient stress field distribution map; the defect prediction and early warning module is used to receive multi-dimensional data streams to generate the probability of defect occurrence. The early warning generation module is used to generate graded early warning signals based on the probability of defect occurrence and preset graded early warning thresholds; The adaptive control module is used to calculate the adjusted pulse energy parameters based on the total interface stress and the graded early warning signal, using a pulse energy control law. The laser processing execution module is used to receive the adjusted pulse energy parameters to achieve closed-loop control of the laser processing process.

Images