Electronic cigarette internal structure multi-mode defect detection method based on intelligent endoscope

Abstract

The invention discloses an electronic cigarette internal structure multi-mode defect detection method based on an intelligent endoscope, and relates to the technical field of artificial intelligence, and the method comprises the steps: constructing an electronic cigarette multi-channel intelligent endoscope system; deploying a physically constrained defect evolution model on the edge computing unit; dynamically generating an optimal detection strategy set by utilizing a migration reinforcement learning algorithm based on the prediction map and the current working mode constraint; when the defect confidence exceeds a dynamic threshold value or the critical failure time is lower than a safety margin, triggering an adaptive focusing controller to execute high-precision reinspection, and activating a miniature printing module accompanying defect verification in the endoscope system; and comparing and predicting the deviation between the defect evolution trajectory and the actual detection result in real time. According to the electronic cigarette internal structure multi-mode defect detection method based on the intelligent endoscope, mechanical defects of traditional fixed detection regulations are overcome, and accurate adaptation of detection resources and risk levels is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an electronic cigarette internal structure multi-modal defect detection method based on an intelligent endoscope. BACKGROUND

[0002] Current electronic cigarette internal defect detection mainly relies on single optical endoscope or offline X-ray scanning, which has significant limitations: traditional optical imaging is easily disturbed by condensate reflection and misses micro-cracks, and cannot perceive thermal stress abnormalities; X-ray detection needs to disassemble the atomization chamber, which destroys the production line continuity. Existing methods can only identify static defects and lack the ability to predict defect evolution paths, resulting in hidden failures causing product failure without early warning.

[0003] In recent years, although some research has attempted to introduce deep learning algorithms, there are still key bottlenecks: first, multi-modal data is time and space dislocated due to endoscope movement, and feature fusion credibility is low; second, data-driven models ignore material thermodynamic constraints and mistakenly identify normal temperature fluctuations as defects; third, detection strategies rely on fixed rules and cannot dynamically adapt to different electronic cigarette structures and working modes. In addition, virtual defect prediction results lack entity verification methods, which restricts the closed-loop evolution ability of the detection system. SUMMARY

[0004] In order to overcome the deficiencies of the prior art, the present application provides an electronic cigarette internal structure multi-modal defect detection method based on an intelligent endoscope.

[0005] In a first aspect, the present application provides an electronic cigarette internal structure multi-modal defect detection method based on an intelligent endoscope, the method comprising: An electronic cigarette multi-channel intelligent endoscope system is constructed, integrating a polarized light imaging module, a macro thermal imaging module, and a laser displacement sensing module, to synchronously obtain multi-modal data streams of the internal structure of the atomization chamber in real time, including surface micro-crack polarization feature maps, heater temperature field distribution, and three-dimensional deformation topology; A physically constrained defect evolution model is deployed on an edge computing unit, and through a cross-modal feature fusion network, the defect coupling effects of different modalities are associated to generate a prediction atlas containing defect type probability matrix, defect propagation path, and critical failure time; Based on the prediction atlas and the current working mode constraints, a transfer reinforcement learning algorithm is used to dynamically generate an optimal detection strategy set, including endoscope pose adjustment vectors, multi-modal sampling weight distribution schemes, and defect verification priorities; When the defect confidence exceeds the dynamic threshold or the critical failure time is lower than the safety margin, an adaptive focusing controller is triggered to perform high-precision re-inspection, and a micro-printing module for accompanying defect verification is activated in the endoscope system; Real-time comparison of the predicted defect evolution trajectory and the actual detection results deviation, through the adversarial domain adaptation training iterative update defect evolution model parameters, form a defect perception-decision-validation closed-loop evolution system.

[0006] Preferably, a micro-prism array with a diameter less than a threshold value is used to realize lateral imaging, solving the problem of detection in the blind area of the long and narrow cavity of the atomizing chamber; Embedding a liquid crystal adjustable phase retarder in the polarized light imaging module can dynamically switch the polarization direction to enhance the crack contrast of different material surfaces; A laser displacement sensing module emits a preset wavelength of fan-shaped structured light, and a Moire fringe analysis method is used to reconstruct the three-dimensional deformation of the ceramic matrix micropores of the atomizing core; A multi-modal space-time alignment mechanism is designed, and the heating wire working current pulse is used as a synchronous trigger signal to eliminate the space-time misalignment between modalities caused by the movement of the endoscope.

[0007] Preferably, a collaborative decision-making framework is established with power reliability agents, energy efficiency optimization agents, and emission control agents as the core, and each agent explores strategies based on independent reward functions; A multi-physical field constraint framework is established with material thermal stress equations, fluid surface tension models, and electro-thermal coupling effects as the core; A double-branch cross-modal feature fusion network is constructed, the first branch uses a graph attention mechanism to extract abnormal heat spot diffusion patterns in thermal imaging temperature fields, and the second branch uses a topological data analysis algorithm to identify stress concentration areas in three-dimensional deformations; The multi-physical field constraints are embedded in the network loss function in the form of partial differential equation residual terms, and the ceramic micro-crack propagation rate, the oil guide cotton fiber swelling degree, and the electrode oxide layer peeling trend are predicted simultaneously. Through defect path cellular automata deduction, the cross-component transmission path of defects under different working temperature / power combinations is simulated.

[0008] Preferably, in the switching preparation stage, according to the high probability fault point position of the prediction map, the ship power grid topology is dynamically reconstructed to isolate the weak nodes; A virtual negative impedance is preset on the auxiliary power side through an impedance reshaping solid-state circuit breaker to offset the high-frequency oscillation risk in the prediction; A power distribution algorithm based on quantum particle swarm optimization is used to dynamically distribute compensation power among supercapacitors, flywheel energy storage, and fuel cells, ensuring that the bus voltage fluctuation rate is less than a preset value; An active damping current is injected at the power switching moment, and its phase is opposite to that of the predicted harmonic spectrum, realizing feedforward cancellation of fault oscillation.

[0009] Preferably, a collaborative decision-making architecture is established to detect efficiency agents, defect coverage agents, and resource consumption agents. The detection efficiency agent reward function includes the defect detection rate per unit time and the attitude adjustment energy consumption coefficient. The defect coverage agent reward function includes the multi-modal feature complementary gain and the key hidden area exploration integrity. The resource consumption agent reward function includes the calculation load balancing degree and the printing material loss cost. A course transfer learning strategy is adopted to distill the detection strategy knowledge of historical electronic cigarette models to the new model decision network. A Pareto optimal strategy set is generated through Monte Carlo tree search, and millisecond-level virtual verification is performed in the digital twin.

[0010] Preferably, according to the defect type probability matrix, the detection modal combination is dynamically switched, the polarization light and laser displacement dual-mode super-resolution scanning is started for micro-crack defects, and the thermal imaging and electrochemical impedance spectroscopy joint diagnosis is started for electrode oxidation defects. The endoscope objective focal length is adjusted by a piezoelectric ceramic micro-motion mechanism to achieve a preset level of focusing accuracy. A thermosetting photosensitive resin is loaded in the micro-printing module, and a micro- entity model is printed according to the predicted defect expansion path for multi-modal cross-validation.

[0011] Preferably, an explainable decision traceability engine is constructed: when a critical defect is detected, a three-dimensional visualization report is automatically generated, and a physical field simulation cloud map, defect evolution path deduction, and multi-agent decision weight distribution are superimposed and displayed. An adversarial sample protection mechanism is designed: a random resonance filter is added at the front end of the feature extraction to suppress feature confusion caused by endoscope metal reflection. A cloud-edge collaborative evolution framework is established: multi-line detection data is aggregated through federated learning to update the shared base model, and the model is compressed to the edge device using knowledge distillation technology.

[0012] Preferably, the explainable neurons in the cross-modal feature fusion network that meet the multi-physical field constraints are extracted to generate a simplified form of the defect thermodynamic evolution equation. An counterfactual reasoning technology is used to simulate the failure scenario when a defect is not detected, and the expected service life decay value of the key components is quantitatively displayed. In the augmented reality interface, the holographic image inside the atomizing cartridge is projected, the defect position is dynamically marked, and the material stress cloud map is associated.

[0013] Preferably, the detection parameters are dynamically adjusted according to the electronic cigarette working mode, a high-frequency periodic scanning strategy is used in the constant power output mode, and an intermittent detection strategy triggered by the heating cycle is used in the pulse mode. The working condition parameters, material batch, and environmental stress factors of historical defect cases are associated to generate a set of defect root cause reasoning rules.

[0014] Preferably, a fuzzy control rule table is constructed with the viscosity coefficient of tobacco tar, the working voltage fluctuation rate and the environmental temperature and humidity as input variables; The missed detection risk curve of different detection strategies is predicted by the long short-term memory network; The Pareto frontier of detection accuracy and resource consumption is solved by the Bayesian optimization algorithm, and the optimal detection parameter combination is output.

[0015] Compared with the prior art, the present application has the following characteristics and beneficial effects: Firstly, by constructing a multi-channel intelligent endoscope system integrating polarized light, thermal imaging and laser displacement sensing, the synchronous capture of internal surface micro-cracks, temperature field distribution and three-dimensional deformation of the atomizing chamber is realized, solving the visual field blind area and data fragmentation problem of traditional single-mode detection, and providing a global perception basis for hidden defect identification. Secondly, the physical constraint defect evolution model is deployed at the edge, and the physical laws such as material thermal stress equation are embedded into the cross-modal feature fusion network, breaking through the empirical limitations of pure data-driven algorithms, generating a prediction atlas containing defect type probability, expansion path and failure time at the same time, and significantly improving the physical rationality and timing accuracy of defect evolution trend prediction. Further, based on the prediction atlas and working constraints, a detection strategy set is dynamically generated using transfer reinforcement learning, and through multi-agent collaborative optimization of endoscope motion trajectory, modal weight allocation and verification priority, the mechanical defects of traditional fixed detection procedures are overcome, and precise adaptation of detection resources and risk levels is realized. In the execution phase, combined with adaptive micrometer-level focusing and micro-3D printing entity verification technology, the multi-modal combination is intelligently switched according to the defect type and a physical verification model is generated, effectively suppressing the false alarm risk caused by image noise in traditional detection, and ensuring zero-miss judgment of key defects. Finally, through real-time comparison of the predicted trajectory and the actual detection result, the defect evolution model parameters are continuously iterated using the adversarial training mechanism, so that the system can adapt to new electronic cigarette structures and unknown defect modes, forming a closed-loop intelligent system with dynamic evolution of detection capability. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a step block diagram of an electronic cigarette internal structure multi-modal defect detection method based on an intelligent endoscope mainly embodied by the present embodiment. DETAILED DESCRIPTION

[0017] The present application will be further described in detail below in conjunction with the following embodiments.

[0018] Reference Figure 1 , the electronic cigarette internal structure multi-modal defect detection method based on an intelligent endoscope, the method comprising the following steps: S1. Constructing an electronic cigarette multi-channel intelligent endoscope system, integrating a polarized light imaging module, a macro thermal imaging module, and a laser displacement sensing module, real-time synchronization of multi-modal data streams of the internal structure of the atomization chamber, including surface micro-crack polarization feature map, heating element temperature field distribution, and three-dimensional deformation topology.

[0019] S2. Deploying a physically constrained defect evolution model on the edge computing unit, associating the defect coupling effect of different modalities through a cross-modal feature fusion network, generating a prediction atlas containing defect type probability matrix, defect propagation path, and critical failure time.

[0020] S3. Based on the prediction atlas and the current working mode constraint, using a transfer reinforcement learning algorithm to dynamically generate an optimal detection strategy set, including endoscope pose adjustment vector, multi-modal sampling weight distribution scheme, and defect verification priority.

[0021] S4. When the defect confidence exceeds the dynamic threshold or the critical failure time is lower than the safety margin, trigger the adaptive focusing controller to perform high-precision re-inspection, and activate the micro-printing module in the endoscope system for accompanying defect verification.

[0022] S5. Real-time comparison of predicted defect evolution trajectory and actual detection result deviation, iterative update of defect evolution model parameters through adversarial domain adaptation training, forming a closed-loop evolution system of defect perception-decision-verification.

[0023] Firstly, by constructing a multi-channel intelligent endoscope system integrating polarized light, thermal imaging, and laser displacement sensing, the synchronous capture of internal surface micro-cracks, temperature field distribution, and three-dimensional deformation of the atomization chamber is realized, solving the visual blind area and data fragmentation problems of traditional single-mode detection and providing a global perception basis for hidden defect identification. Secondly, by deploying a physically constrained defect evolution model at the edge, the material thermal stress equation and other physical laws are embedded into the cross-modal feature fusion network, breaking through the empirical limitations of pure data-driven algorithms and generating a prediction map that includes defect type probability, expansion path, and failure time, significantly improving the physical reasonableness and timing accuracy of defect evolution trend prediction. Further, based on the prediction map and work constraints, a detection strategy set is dynamically generated using transfer reinforcement learning, and through multi-agent collaborative optimization of endoscope motion trajectory, modal weight allocation, and verification priority, the mechanical defects of traditional fixed detection procedures are overcome, enabling precise adaptation of detection resources and risk levels. In the execution phase, adaptive micrometer-level focusing and micro-3D printing entity verification technology are combined to intelligently switch between multi-modal combinations based on defect type and generate physical verification models, effectively suppressing false alarm risks caused by image noise in traditional detection and ensuring zero-miss judgment of critical defects. Finally, through real-time comparison of predicted trajectories and actual detection results, the defect evolution model parameters are continuously iterated using an adversarial training mechanism, enabling the system to adapt to new electronic cigarette structures and unknown defect patterns, forming a closed-loop intelligent system with dynamic evolution of detection capabilities.

[0024] The specific step S1 includes the following sub-steps: A micro-prism array with a diameter less than a threshold value is used to realize lateral imaging, solving the problem of blind area detection in the narrow cavity of the atomization chamber; A liquid crystal adjustable phase retarder is embedded in the polarized light imaging module to dynamically switch the polarization direction (e.g., 0°, 45°, 90°, 135° in a cycle) to enhance the crack contrast of different materials (such as ceramic matrix and 304 stainless steel electrode); A laser displacement sensing module emits a fan-shaped structured light with a preset wavelength of 650nm, and a Moire fringe analysis method is used to reconstruct the three-dimensional deformation of the ceramic matrix micro-pores of the atomization core, achieving a deformation measurement accuracy of ±2μm; A multi-modal time and space alignment mechanism is designed, using the heating wire working current pulse (pulse width 10ms, frequency 1Hz) as a synchronous trigger signal to ensure that the time synchronization error of the polarized light imaging, thermal imaging, and laser displacement sensing data is less than 1ms, and the spatial registration error is less than 5 pixels, eliminating the time and space misalignment between modalities caused by endoscope movement.

[0025] Specifically, step S1 is a key link in constructing an electronic cigarette multi-channel intelligent endoscope system, aiming to obtain multi-dimensional information inside the atomization chamber in all directions. A micro-prism array with a diameter less than a threshold is used to realize lateral imaging, because the traditional endoscope has difficulty in effectively observing the side wall and other areas of the long and narrow cavity of the atomization chamber, while the micro-prism array can change the direction of light propagation through the principle of optical refraction, thereby solving the problem of blind area detection in the long and narrow cavity of the atomization chamber, and enabling the originally difficult-to-observe areas to be clearly imaged. A liquid crystal adjustable phase retarder is embedded in the polarized light imaging module, and the principle is that the reflection characteristics of polarized light on the surfaces of different materials are different. By dynamically switching the polarization direction, the phase of the polarized light can be adjusted to enhance the crack contrast of different material surfaces. For example, for different materials such as metal and ceramic, under the irradiation of light of a specific polarization direction, the reflected light at the crack will show more obvious differences, thereby more easily identifying surface micro-cracks. The laser displacement sensing module emits a fan-shaped structured light of a predetermined wavelength, and the three-dimensional deformation of the ceramic matrix micro-pores of the atomization core is reconstructed by combining the Moiré fringe analysis method. The fan-shaped structured light can cover a larger detection area, and when the structured light irradiates the surface of the micro-pores, Moiré fringes are formed by interference. By analyzing and calculating these fringes, the three-dimensional shape change of the micro-pores can be accurately measured. A multi-modal space-time alignment mechanism is designed, and the heating wire working current pulse is used as a synchronous trigger signal. This is because the endoscope may move during detection, causing the collection times of different modal data to be out of sync. The heating wire working current pulse has a fixed time interval, which is used as a trigger signal to ensure that the data of different modalities such as polarized light imaging, thermal imaging, and laser displacement sensing are accurately aligned in time and space, eliminating the time and space misalignment between modalities caused by endoscope movement, and ensuring the consistency and reliability of the multi-modal data stream, providing accurate data basis for subsequent cross-modal analysis.

[0026] Construct an endoscope system. Micro-prism array parameter example: use a micro-prism array with a diameter less than 0.5 mm to realize lateral imaging and solve the problem of blind area detection in the long and narrow cavity of the atomization chamber.

[0027] Specific step S2 includes the following sub-steps: Establish a multi-physical field constraint framework with material thermal stress equation (such as Hooke's law combined with thermal expansion coefficient to calculate local stress σ=EαΔT), fluid surface tension model (such as Young-Laplace equation to calculate meniscus curve of tobacco tar in micro-pores), and electro-thermal coupling effect (Joule's law P=I²R to calculate heating body power) as the core; Construct a double-branch cross-modal feature fusion network, the first branch uses an 8-layer graph attention mechanism (GAT) to extract abnormal heat spot diffusion patterns in the thermal imaging temperature field, and the second branch identifies stress concentration areas (Betti number abnormal areas) in three-dimensional deformation through topological data analysis algorithms; The multi-physics constraints are embedded in the network loss function (total loss = cross-entropy + λ‖Res‖²) in the form of partial differential equation residual terms (e.g., the residual of the heat conduction equation ∇·(k∇T)+Q=0, ‖Res‖²), which simultaneously predicts the ceramic micro-crack propagation rate (unit: μm / hour), the oil guiding cotton fiber swelling degree (volume expansion percentage), and the electrode oxide layer peeling trend (oxide layer thickness thinning rate); The defect path cellular automaton is deduced to simulate the cross-component infection path of defects under different working temperature / power combinations.

[0028] Specifically, step S2 focuses on building a high-precision defect evolution model in the edge computing unit. A multi-physical field constraint framework is established, with material thermal stress equation, fluid surface tension model, and electro-thermal coupling effect as the core, because the defect evolution inside the electronic cigarette atomization cartridge is affected by multiple physical factors. The material thermal stress equation can describe the stress distribution and deformation of ceramic, metal and other materials under temperature change, the fluid surface tension model can explain the effect of the behavior of fluid such as tobacco oil on the surface on the defect, and the electro-thermal coupling effect reflects the interaction between the heat generated by the current through the heating body and the electric field. The comprehensive consideration of these physical laws provides a solid theoretical basis for the defect evolution model. A double-branch cross-modal feature fusion network is constructed. The first branch uses a graph attention mechanism to extract abnormal heat spot diffusion patterns in the thermal imaging temperature field. The graph attention mechanism can automatically assign attention weights to different heat spots according to the correlation between heat spots, so as to more accurately identify the diffusion trend of heat spots, which is of great significance for discovering temperature abnormal distribution of the heating body and potential defects. The second branch identifies stress concentration areas in three-dimensional deformation through topological data analysis algorithms. Topological data analysis can extract key topological features such as holes and cracks from complex three-dimensional deformation data, and then determine the location of stress concentration, providing a basis for defect positioning. The multi-physical field constraints are embedded in the network loss function in the form of partial differential equation residual terms. This can force the model to predict results that conform to physical laws during network training. For example, when predicting the expansion rate of ceramic micro-cracks, the residual term will constrain the predicted result to be as close as possible to the theoretical expansion rate calculated by the material thermal stress equation, thereby simultaneously predicting the expansion rate of ceramic micro-cracks, the swelling degree of oil guide cotton fibers, and the oxidation layer peeling trend of the electrode, improving the accuracy and physical reasonableness of the prediction. Through defect path cellular automata deduction, the cell size is set to 50μm x 50μm, and the state transition rules are based on the stress difference of adjacent cells, temperature gradient, and material fatigue limit. The simulation of the cross-component infection path of defects under different working temperature / power combinations, cellular automata can divide the atomization cartridge into multiple cells, each cell evolves according to the state of the surrounding cells and the preset rules, thereby simulating the cross-component infection path of defects (such as ceramic cracks spreading to the electrode) under different temperatures (such as 200°C, 250°C, 300°C) and powers (such as 8W, 12W, 15W). The process of spreading from one component to another provides support for a comprehensive understanding of the propagation rules of defects and the development of effective detection strategies.

[0029] Specific step S3 includes the following sub-steps: A collaborative decision-making architecture is established, which is composed of detection efficiency agent, defect coverage agent and resource consumption agent. The reward function of detection efficiency agent includes defect detection rate per unit time (weight 0.4) and posture adjustment energy consumption coefficient (weight 0.3). The reward function of defect coverage agent includes multi-modal feature complementary gain (such as polarization + thermal imaging combination gain factor 1.8, weight 0.5) and key hidden area exploration integrity (weight 0.4). The reward function of resource consumption agent includes calculation load balancing degree (weight 0.6) and printing material loss cost (weight 0.4). A course transfer learning strategy is adopted to distill the detection strategy knowledge of historical electronic cigarette models to the new model decision network. A Pareto optimal strategy set is generated through Monte Carlo tree search, and millisecond-level virtual verification is performed in the digital twin.

[0030] Specifically, step S3 aims to generate an optimal strategy set that adapts to different detection scenarios. A collaborative decision-making framework is established, consisting of a detection efficiency agent, a defect coverage agent, and a resource consumption agent. The reward function of the detection efficiency agent includes the defect detection rate per unit time and the posture adjustment energy consumption coefficient. The defect detection rate per unit time reflects the speed and efficiency of detection, and the posture adjustment energy consumption coefficient considers the energy consumed by adjusting the posture of the endoscope. By integrating these two factors, the agent is prompted to improve detection efficiency while minimizing energy consumption. The reward function of the defect coverage agent includes multi-modal feature complementary gain and key hidden area exploration integrity. Multi-modal feature complementary gain emphasizes the complementary effect between different modal data, such as the combination of polarized light imaging and thermal imaging data, which can more comprehensively describe defect features. Key hidden area exploration integrity ensures that important hidden areas are not missed during the detection process, thereby improving the comprehensiveness of defect detection. The reward function of the resource consumption agent includes computational load balancing degree and printing material loss cost. Computational load balancing degree ensures that the computational resources of edge computing units are reasonably allocated to avoid overload. Printing material loss cost considers the material consumption of the micro-printing module, achieving efficient use of resources. A curriculum transfer learning strategy is adopted to transfer the detection strategy knowledge of historical electronic cigarette models (such as model A) to the new model (such as model A) decision network. The initial learning rate is set to 0.01 and adjusted dynamically according to the curriculum difficulty (complexity of the new model structure). Curriculum transfer learning can gradually transfer the detection strategy knowledge of historical models to the new model decision network in order of difficulty, such as first learning basic detection strategies and then learning complex adaptive strategies. This can accelerate the training process of the new model decision network and enable it to quickly adapt to new electronic cigarette models. A Pareto optimal strategy set is generated through Monte Carlo tree search (simulation times > 1000) (selecting strategies with detection efficiency > 95%, defect coverage > 92%, and resource consumption < threshold), and virtual verification is performed in a digital twin built based on the Unity engine at a millisecond level (< 10ms / time). Monte Carlo tree search finds the Pareto optimal solution among numerous possible strategies through a large number of simulations and sampling, i.e., strategies that balance detection efficiency, defect coverage, and resource consumption. The digital twin can accurately simulate the actual detection environment of the electronic cigarette atomization chamber, quickly verify these strategies in a virtual environment, evaluate their performance in different scenarios, and thus select the optimal detection strategy set to provide reliable guidance for actual detection.

[0031] Specific step S4 includes the following sub-steps: According to the defect type probability matrix, dynamically switch the detection modal combination, start polarized light and laser displacement dual-modal super-resolution scanning for micro-crack defects, and start thermal imaging and electrochemical impedance spectroscopy joint diagnosis for electrode oxidation defects. The focal length of the endoscope objective is adjusted by a piezoelectric ceramic micro-motion mechanism to achieve preset level focusing accuracy. In the micro-printing module, thermosetting photosensitive resin is loaded to print micro-physical models according to the predicted defect propagation path for multi-modal cross-validation.

[0032] Specifically, step S4 focuses on high-precision re-inspection and verification of suspected defects. According to the defect type probability matrix, the detection modal combination is dynamically switched. For micro-crack defects (probability > 85%), polarization light (resolution improved to 2 μm / pixel) and laser displacement (scanning density improved to 0.1 mm interval) dual-modal super-resolution scanning are started. For electrode oxidation defects (probability > 80%), thermal imaging (temperature sensitivity 0.1 °C) and electrochemical impedance spectroscopy (frequency range 1 kHz-100 kHz) joint diagnosis is started.

[0033] This is because polarization light imaging has high sensitivity to surface micro-cracks and can clearly show the morphology and distribution of cracks, while laser displacement sensing can provide three-dimensional position information of micro-cracks. The dual-modal super-resolution scanning combined with the two can achieve high-precision detection of micro-cracks and improve the accuracy and resolution of detection. For electrode oxidation defects, thermal imaging and electrochemical impedance spectroscopy joint diagnosis is started. Thermal imaging can detect temperature distribution anomalies of the electrode, reflecting the oxidation degree of the electrode, and electrochemical impedance spectroscopy can measure the change of the electrical properties of the electrode. The combination of the two can more comprehensively evaluate the severity of electrode oxidation defects. The focal length of the endoscope objective is adjusted by a piezoelectric ceramic micro-motion mechanism to achieve preset level focusing accuracy. Piezoelectric ceramics have high precision displacement control capability and can accurately adjust the position of the endoscope objective, thereby achieving micron-level or even higher precision focusing, ensuring that the details of the defect can be clearly observed during re-inspection. In the micro-printing module, thermosetting photosensitive resin is loaded to print micro-physical models according to the predicted defect propagation path for multi-modal cross-validation. Thermosetting photosensitive resin will quickly solidify under specific light conditions, and can accurately print micro-physical models consistent with the predicted defect propagation path. By detecting the physical models with multiple modalities such as polarization light imaging and thermal imaging, the original detection data can be compared and verified, effectively suppressing the false alarm risk caused by image noise in traditional detection, ensuring zero-miss judgment of critical defects, and improving the reliability of the detection results.

[0034] The specific step S5 process can be that step S5 is the core step of forming a defect detection closed-loop evolution system. In real-time comparison of the predicted defect evolution trajectory (such as the crack length prediction value L_pred) and the actual detection result deviation (actual value L_real), the defect evolution model parameters are iteratively updated through the adversarial domain adaptation training based on the Wasserstein distance (the generator and the discriminator are updated alternately), and the iteration period is set to update once every 1000 products detected, forming a closed-loop evolution system of defect perception-decision-verification.

[0035] By collecting the actually detected defect data and comparing it with the previously predicted evolution trajectory, the difference between the two is calculated. This deviation analysis can reveal the shortcomings of the model in the prediction process, such as the possibility of not fully considering some physical factors or not accurately extracting features. Through the iterative updating of the defect evolution model parameters by the adversarial domain adaptation training, the core idea of the adversarial domain adaptation training is to enable the model to adapt to different detection environments and defect types. By constructing an adversarial network, the model can quickly adjust its parameters when facing new and unseen defect data, improving the accuracy of prediction. For example, when a new type of micro-crack defect is detected, the adversarial domain adaptation training can prompt the model to adjust its internal feature extraction and prediction algorithm, enabling it to better recognize this new type of defect. Through continuous iteration and updating, a closed-loop evolution system of defect perception-decision-verification is formed, in which the defect perception module is responsible for collecting and analyzing multi-modal data, the decision module generates detection strategies based on the perceived information, and the verification module verifies and feeds back the detection results. The three modules cooperate with each other to continuously optimize the performance of the entire detection system, enabling it to continuously adapt to new electronic cigarette structures and unknown defect patterns, improving detection capability and accuracy.

[0036] In some embodiments, the intelligent endoscope-based multi-modal defect detection method for electronic cigarette internal structure can further include the following steps: Build an explainable decision traceability engine: when a critical defect is detected, automatically generate a three-dimensional visualization report, superimposed with physical field simulation cloud maps, defect evolution path deduction, and multi-agent decision weight distribution; Design an adversarial sample protection mechanism: add a random resonance filter at the front end of feature extraction to suppress feature confusion caused by endoscope metal reflections; Establish a cloud-edge collaborative evolution framework: aggregate multi-line detection data to update a shared base model through federated learning, and compress the model to edge devices using knowledge distillation technology.

[0037] Specifically, the explainable decision trace engine is built to improve the transparency and reliability of the detection system. When a critical defect is detected, a three-dimensional visualization report is automatically generated, superimposed with physical field simulation cloud maps, defect evolution path deduction, and multi-agent decision weight distribution. The physical field simulation cloud map can intuitively show the distribution of physical fields around the defect, such as temperature field, stress field, etc., helping technicians better understand the formation environment of the defect. Defect evolution path deduction simulates the whole process from defect generation to development, enabling technicians to understand the evolution law of the defect. The multi-agent decision weight distribution shows the contribution of each agent in the decision-making process, facilitating the analysis of the rationality of the decision. The design of the adversarial sample protection mechanism is to deal with interference factors in the detection process. A stochastic resonance filter is added at the front end of feature extraction, which can enhance useful signals and suppress feature confusion caused by endoscope metal reflection. For example, when the endoscope detects metal parts, metal reflection may produce strong noise signals, and the stochastic resonance filter can convert these noise signals into useful information, improving the accuracy of feature extraction. The cloud-edge collaborative evolution framework is established to realize the collaborative optimization of the detection system. Through federated learning, multiple production line detection data are aggregated to update the shared base model, and federated learning allows different production lines to train models without sharing raw data, protecting data privacy. Knowledge distillation technology is used to compress the model to the edge device, which can compress complex models into lightweight models, facilitating deployment and operation on edge devices, improving the real-time performance and response speed of the detection system.

[0038] In some embodiments, the method for detecting internal structural multi-modal defects of an intelligent endoscope-based electronic cigarette can further include the following steps: Extracting explainable neurons in the cross-modal feature fusion network that meet the multi-physical field constraints to generate a simplified form of the defect thermodynamic evolution equation; Using counterfactual reasoning technology to simulate failure scenarios when defects are not detected, and quantitatively displaying the expected service life decay value of key components; Projecting holographic images inside the atomization chamber in the augmented reality interface, dynamically marking defect locations and correlating material stress cloud maps.

[0039] Specifically, the interpretable neurons in the cross-modal feature fusion network that satisfy the multi-physical field constraints are extracted, and a simplified form of the defect thermodynamic evolution equation is generated, which helps to deeply understand the physical mechanism of defect evolution. The interpretable neurons refer to those neurons in the network that are sensitive to physical field constraints. By analyzing these neurons, key features related to defect thermodynamic evolution can be extracted, and a simplified evolution equation can be derived, providing a more concise and accurate mathematical model for defect prediction and analysis. The counterfactual reasoning technique is used to simulate failure scenarios when defects are not detected, and the expected life degradation value of the key components is quantified. Counterfactual reasoning analyzes the impact of defects on the life of key components by assuming different detection results. For example, assuming that a certain micro-crack defect is not detected, the development of the defect during subsequent use is simulated, and the expected life degradation value of the key component is calculated, which is of great significance for evaluating the reliability and safety of the product. In the augmented reality interface, the holographic image inside the atomizing chamber is projected, the defect position is dynamically marked, and the material stress cloud map is associated. Augmented reality technology can combine virtual defect information with the actual internal structure of the atomizing chamber, providing an intuitive and three-dimensional display of the detection results for technicians. The material stress cloud map can show the stress distribution around the defect, helping technicians better understand the mechanical properties and potential risks of the defect.

[0040] In some embodiments, the method for multi-modal defect detection of the internal structure of the smart endoscope-based electronic cigarette can further include the following steps: Adjust the detection parameters dynamically according to the working mode of the electronic cigarette, use high-frequency periodic scanning strategy in constant power output mode, and use intermittent detection strategy triggered by heating cycle in pulse mode; Correlate the working condition parameters, material batch, and environmental stress factors of historical defect cases to generate a set of defect root cause reasoning rules.

[0041] Specifically, the detection parameters are adjusted dynamically according to the working mode of the electronic cigarette, and high-frequency periodic scanning strategy is used in constant power output mode. In constant power output mode, the working state of the electronic cigarette is relatively stable, and high-frequency periodic scanning can quickly detect possible defects, improving detection efficiency. In pulse mode, intermittent detection triggered by heating cycle is used. In pulse mode, the working state of the electronic cigarette changes frequently, and intermittent detection triggered by heating cycle can reduce the number of detections and resource consumption without affecting the detection effect. Correlate the working condition parameters, material batch, and environmental stress factors of historical defect cases to generate a set of defect root cause reasoning rules. Through analysis of historical defect cases, key factors leading to defects are found, such as specific working condition parameters, material batch, or environmental stress, and a set of reasoning rules is established. When a new defect is detected, the root cause of the defect can be quickly inferred according to these rules, providing guidance for defect repair and prevention.

[0042] In some embodiments, the intelligent endoscope-based electronic cigarette internal structure multi-modal defect detection method can further include the following steps: A fuzzy control rule table is constructed with the viscosity coefficient of the tobacco tar, the working voltage fluctuation rate and the environmental temperature and humidity as input variables. The long short-term memory network is used to predict the missed detection risk curve of different detection strategies. The Bayesian optimization algorithm is used to solve the Pareto front of detection accuracy and resource consumption, and output the optimal detection parameter combination.

[0043] Specifically, a fuzzy control rule table is constructed with the viscosity coefficient of the tobacco tar, the working voltage fluctuation rate and the environmental temperature and humidity as input variables. The viscosity coefficient of the tobacco tar, the working voltage fluctuation rate and the environmental temperature and humidity and other factors will affect the defects of the electronic cigarette. The fuzzy control rule table fuzzily processes the relationship between these factors and the detection parameters, which can more flexibly cope with complex detection environments. For example, when the viscosity coefficient of the tobacco tar is high, the frequency and accuracy of detection may need to be adjusted. The long short-term memory network is used to predict the missed detection risk curve of different detection strategies. The long short-term memory network has strong time series data processing capability and can predict the missed detection risk under different detection strategies according to historical detection data. For example, for a specific detection strategy, the long short-term memory network can predict the missed detection risk change at different time points, providing a basis for selecting the optimal detection strategy. The Bayesian optimization algorithm is used to solve the Pareto front of detection accuracy and resource consumption, and output the optimal detection parameter combination. The Bayesian optimization algorithm can efficiently find the optimal solution in an uncertain search space by balancing detection accuracy and resource consumption, finding the best balance point between the two, i.e. the Pareto front, thereby outputting the optimal detection parameter combination and achieving efficient operation of the detection system.

[0044] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Therefore, any equivalent changes made in terms of structure, shape and principle based on the present application should be covered within the protection scope of the present application.

Claims

1. A smart endoscope-based multi-modal defect detection method for internal structure of an electronic cigarette, characterized in that, The method comprises the following steps: An electronic cigarette multi-channel intelligent endoscope system is constructed, which integrates a polarized light imaging module, a macro thermal imaging module and a laser displacement sensing module, and real-time and synchronous multi-modal data streams of the internal structure of the atomization chamber are acquired, including a surface micro-crack polarization feature map, a heating element temperature field distribution and a three-dimensional deformation topology; A physically constrained defect evolution model is deployed on an edge computing unit, defect coupling effects of different modalities are associated through a cross-modal feature fusion network, and a prediction graph containing a defect type probability matrix, a defect propagation path and a critical failure time is generated; Based on the prediction graph and the current working mode constraint, a transfer reinforcement learning algorithm is used to dynamically generate an optimal detection strategy set, including an endoscope posture adjustment vector, a multi-modal sampling weight distribution scheme and a defect verification priority; When the defect confidence exceeds a dynamic threshold or the critical failure time is lower than a safety margin, an adaptive focusing controller is triggered to perform high-precision re-inspection, and a micro-printing module for accompanying defect verification is activated in the endoscope system; Real-time deviation of the predicted defect evolution trajectory and the actual detection result is compared, and the defect evolution model parameters are iteratively updated through adversarial domain adaptation training, forming a closed-loop evolution system of defect perception, decision-making and verification.

2. The method for multimodal defect detection of the internal structure of an electronic cigarette based on an intelligent endoscope according to claim 1, characterized in that: An electronic cigarette multi-channel intelligent endoscope system is constructed, which integrates a polarized light imaging module, a macro thermal imaging module and a laser displacement sensing module, and real-time and synchronous multi-modal data streams of the internal structure of the atomization chamber are acquired, including a surface micro-crack polarization feature map, a heating element temperature field distribution and a three-dimensional deformation topology, and the steps are as follows: A micro-prism array with a diameter less than a threshold value is used to realize lateral imaging, solving the problem of blind area detection in the long and narrow cavity of the atomization chamber; A liquid crystal adjustable phase retarder is embedded in the polarized light imaging module to dynamically switch the polarization direction to enhance the crack contrast of different material surfaces; A laser displacement sensing module emits a fan-shaped structured light of a preset wavelength, and a three-dimensional deformation of the micro-pores of the ceramic substrate of the atomization core is reconstructed by combining the moire fringe analysis method; A multi-modal time-space alignment mechanism is designed, and the heating wire working current pulse is used as a synchronous trigger signal to eliminate the time-space misalignment between modalities caused by endoscope movement.

3. The method for multimodal defect detection of the internal structure of an electronic cigarette based on an intelligent endoscope according to claim 2, characterized in that: A physically constrained defect evolution model is deployed on an edge computing unit, defect coupling effects of different modalities are associated through a cross-modal feature fusion network, and a prediction graph containing a defect type probability matrix, a defect propagation path and a critical failure time is generated, and the steps are as follows: A multi-physical field constraint framework is established, taking material thermal stress equation, fluid surface tension model and electro-thermal coupling effect as the core; A double-branch cross-modal feature fusion network is constructed, the first branch uses a graph attention mechanism to extract abnormal heat spot diffusion patterns in the thermal imaging temperature field, and the second branch uses a topological data analysis algorithm to identify stress concentration areas in the three-dimensional deformation; The multi-physical field constraint is embedded in the network loss function in the form of partial differential equation residual term, and the ceramic micro-crack propagation rate, the oil guide cotton fiber swelling degree and the electrode oxide layer peeling trend are simultaneously predicted; Through defect path cellular automata deduction, the cross-component infection path of defects under different working temperature / power combinations is simulated.

4. The method of claim 3, wherein the method is characterized by: Based on the predicted atlas and the current working mode constraints, an optimal detection strategy set is dynamically generated using a transfer reinforcement learning algorithm, which includes steps of endoscope posture adjustment vector, multi-modal sampling weight distribution scheme and defect verification priority, specifically: A collaborative decision-making framework is established, which is composed of detection efficiency agents, defect coverage agents and resource consumption agents. The reward function of the detection efficiency agent includes the defect detection rate per unit time and the posture adjustment energy consumption coefficient. The reward function of the defect coverage agent includes the multi-modal feature complementary gain and the key hidden area exploration integrity. The reward function of the resource consumption agent includes the calculation load balancing degree and the printing material loss cost. A curriculum transfer learning strategy is adopted to distill the detection strategy knowledge of historical electronic cigarette models to the new model decision network. A Pareto optimal strategy set is generated through Monte Carlo tree search, and millisecond-level virtual verification is performed in the digital twin.

5. The method for multimodal defect detection of the internal structure of an electronic cigarette based on an intelligent endoscope according to claim 4, characterized in that: When the defect confidence exceeds the dynamic threshold or the critical failure time is lower than the safety margin, an adaptive focusing controller is triggered to perform high-precision re-inspection, and a micro-printing module is activated for defect verification in the endoscope system. According to the defect type probability matrix, the detection modal combination is dynamically switched. Polarized light and laser displacement dual-mode super-resolution scanning are started for micro-crack defects, and thermal imaging and electrochemical impedance spectroscopy are started for electrode oxidation defects. The focal length of the endoscope objective lens is adjusted by a piezoelectric ceramic micro-motion mechanism to achieve a preset level of focusing accuracy. A thermosetting photosensitive resin is loaded in the micro-printing module, and a micro-observation entity model is printed according to the predicted defect expansion path for multi-modal cross-validation.

6. The method of claim 1, wherein the method is a method of multi-modal defect detection of an internal structure of an electronic cigarette based on a smart endoscope, the method comprising: The method further comprises: An explainable decision traceability engine is constructed: when a critical defect is detected, a three-dimensional visual report is automatically generated, and physical field simulation cloud maps, defect evolution path deduction and multi-agent decision weight distribution are superimposed and displayed; An adversarial sample protection mechanism is designed: a random resonance filter is added at the front end of the feature extraction to suppress feature confusion caused by endoscope metal reflection; A cloud-edge collaborative evolution framework is established: multi-line detection data is aggregated through federated learning to update the shared base model, and the model is compressed to the edge device using knowledge distillation technology.

7. The method for multimodal defect detection of the internal structure of an electronic cigarette based on an intelligent endoscope according to claim 6, characterized in that: An explainable decision traceability engine is constructed: when a critical defect is detected, a three-dimensional visual report is automatically generated, and physical field simulation cloud maps, defect evolution path deduction and multi-agent decision weight distribution are superimposed and displayed. Explainable neurons that meet the multi-physical field constraints in the cross-modal feature fusion network are extracted to generate a simplified form of the defect thermodynamic evolution equation; An counterfactual reasoning technology is used to simulate the failure scenario when the defect is not detected, and the expected service life decay value of the key components is quantitatively displayed; In the augmented reality interface, the holographic image inside the atomizing cartridge is projected, and the defect position is dynamically marked and associated with the material stress cloud map.

8. The method of claim 1, wherein the method is a method of intelligent endoscope-based multi-modal defect detection of internal structure of an electronic cigarette, characterized in that, The method further comprises: Detection parameters are dynamically adjusted according to the working mode of the electronic cigarette. In the constant power output mode, a high-frequency periodic scanning strategy is adopted, and in the pulse mode, an intermittent detection strategy triggered according to the heating cycle is adopted. The working condition parameters, material batch and environmental stress factors of historical defect cases are associated to generate a set of defect root cause reasoning rules.

9. The method of claim 8, wherein the method further comprises: According to the electronic cigarette working mode dynamic adjustment detection parameter, in the constant power output mode adopts high frequency cycle scanning strategy, in the pulse mode adopts the step of intermittent detection strategy triggered by heating cycle, specifically: A fuzzy control rule table is constructed with the viscosity coefficient of tobacco tar, the working voltage fluctuation rate and the environmental temperature and humidity as input variables; The long short-term memory network is used to predict the missed detection risk curve of different detection strategies; The Bayesian optimization algorithm is used to solve the Pareto front of detection accuracy and resource consumption, and the optimal detection parameter combination is output.

Images