A film air tightness detection device of a microfluidic chip

Abstract

This invention relates to the field of airtightness testing technology, and more particularly to a thin-film airtightness testing device for microfluidic chips. This device generates a spatial leakage probability map through multi-frequency joint scanning for initial screening, identifying overt leaks and hidden defect areas. Fractal pressure waves are applied to the hidden areas, and the detection sensitivity is dynamically adjusted based on signal attenuation characteristics to extract the defect response intensity. A long short-term memory network prediction model is constructed based on the time-series data of the response intensity to quantify the defect evolution risk and make intelligent decisions. Combined with a pre-embedded microcapsule system, it enables timely repair of overt leaks and preventative intervention for high-risk hidden defects. This invention achieves high-sensitivity detection of micro / nano-level defects, dynamic prediction of defect evolution risk, and precise automated repair of leak points, thereby constructing a closed-loop intelligent maintenance system integrating detection, judgment, prediction, repair, and self-optimization, significantly improving the long-term reliability and service life of microfluidic chips.

Description

Technical Field

[0001] This invention relates to the field of airtightness testing technology, and more particularly to a thin-film airtightness testing device for a microfluidic chip. Background Technology

[0002] The thin-film airtightness testing of microfluidic chips is crucial for ensuring chip functionality and reliability. However, current testing technologies have long relied on traditional methods such as water detection and pressure drop, which suffer from insufficient sensitivity, easy contamination or damage during testing, and difficulty in quantification. These methods cannot meet the production and R&D needs for quantitative testing of microfluidic chips.

[0003] Chinese Patent Publication No. CN112504564A discloses a device and method for detecting the airtightness of microfluidic chips after film bonding. The related technical solution includes horizontally transporting the test piece to a testing platform using a feeding suction cup, a feeding cylinder, and a rodless cylinder; receiving and fixing the test chip on the testing platform, and driving a pressure head downward to fill the sealed cavity with testing gas; after the test is completed, the unloading suction cup picks up the chip according to the test results, and the unloading cylinder sorts the chip to the unloading platform in the qualified or unqualified area, thus completing the unloading process. However, the related technical solution is an automated packaging method based on the traditional pressure attenuation method, which lacks high sensitivity detection capability for minute leaks, does not integrate a real-time temperature compensation mechanism, and cannot output spatial positioning information, thereby affecting the final product's yield and long-term reliability.

[0004] Therefore, there is an urgent need for a thin-film airtightness detection device that can achieve high sensitivity, location, online monitoring and intelligent early warning functions for microfluidic chips, and repair leaks or defects through the device, thereby improving the yield and long-term reliability of the final product. Summary of the Invention

[0005] To address these issues, the present invention provides a thin-film airtightness detection device for microfluidic chips, which overcomes the problems of existing technologies that rely on automated packaging of traditional pressure decay methods, lack of high-sensitivity detection capability for minute leaks, lack of integrated real-time temperature compensation mechanism, and inability to output spatial positioning information.

[0006] To achieve the above objectives, the present invention provides a thin-film hermeticity detection device for a microfluidic chip, comprising: The multi-frequency joint scanning module is used to apply multi-band pressure excitation to the thin film region of the microfluidic chip and simultaneously acquire deformation response signals through a capacitive sensing array to generate a spatial leakage probability map. The primary determination module is used to identify regions in the spatial leakage probability map where the signal strength is greater than or equal to the signal strength threshold, determine them as explicit leakage regions and generate a first repair instruction containing the coordinates of the explicit leakage regions, and identify regions where the signal strength is less than the signal strength threshold and the noise level is greater than or equal to the noise level threshold, determine them as hidden defect regions and output the coordinates and signal attenuation characteristics of the hidden defect regions. The fractal excitation module is used to apply fractal pressure waves to the defect concealment area and dynamically extract the signal attenuation characteristics to determine the corresponding attenuation rate. If the attenuation rate is greater than the preset attenuation rate, the response intensity threshold for extracting the defect feature signal is reduced and the defect response intensity is output based on the reduced threshold. The secondary judgment module is used to identify areas where the defect response intensity is greater than or equal to the preset defect intensity, and to determine that there is a high-risk hidden defect. The timing decision module is used to construct a timing prediction model that quantifies the long-term behavior of the defect based on the defect response intensity when the existence of the high-risk hidden defect is determined, and to generate a second repair instruction containing the coordinates of the high-risk defect by combining the pre-stored frequency-current mapping relationship when the predicted risk value output by the timing prediction model is greater than or equal to the intervention risk threshold. The repair execution module is used to locate the coordinates of the explicit leakage area according to the first repair instruction and the coordinates of the high-risk defect according to the second repair instruction, and activate the microcapsule system pre-embedded in the corresponding coordinates.

[0007] Furthermore, the primary determination module includes: The overt leakage processing unit is used to mark the area where the signal strength is greater than or equal to the signal strength threshold as an overt leakage area, extract the geometric center coordinates as output coordinates, generate the first repair instruction, and activate the pre-embedded microcapsule system for timely repair based on the first repair instruction. The hidden defect screening unit is used to mark regions where the signal strength is less than the signal strength threshold and the noise level is greater than or equal to the noise level threshold as hidden defect regions, extract the geometric center coordinates as output coordinates, and calculate the signal attenuation gradient of the hidden defect region in multiple frequency bands as the signal attenuation feature.

[0008] Furthermore, the fractal excitation module includes: A waveform generation unit is used to generate the fractal pressure wave having self-similar spectral characteristics; A response analysis unit is used to perform time-frequency analysis on the current or vibration signal acquired after the fractal pressure wave is applied, in order to determine the attenuation rate of the calculated signal envelope; A threshold adaptive unit is used to reduce the response intensity threshold by a preset ratio when the attenuation rate is greater than the preset attenuation rate.

[0009] Furthermore, the fractal excitation module is also used to determine the direct output of the defect response intensity based on the attenuation rate being less than or equal to the preset attenuation rate.

[0010] Furthermore, the timing decision module includes: The model building unit is used to construct the time-series prediction model based on the time series of the defect response intensity, the corresponding excitation frequency, and the signal attenuation characteristics as inputs, using a long short-term memory network. The risk quantification unit is used to run the time series prediction model and output the predicted risk value; A decision generation unit is used to perform source localization by combining the frequency-current mapping relationship when the predicted risk value is greater than or equal to the intervention risk threshold, so as to generate a second repair instruction containing the coordinates of the high-risk defect.

[0011] Furthermore, the model building unit is used to perform accelerated aging tests on known defect samples and record the corresponding failure times; The model building unit is also used to input the defect state parameters corresponding to a specific percentage safety margin before the failure time into the time-series prediction model, and the output risk value is the intervention risk threshold.

[0012] Furthermore, the repair execution module includes: The coordinate positioning unit is used to receive the coordinates of the explicit leakage area in the first repair instruction and the high-risk coordinates in the second repair instruction, and convert them into addressing signals corresponding to the microcapsule system. An energy excitation unit is used to emit focused ultrasound or a laser of a specific wavelength to the coordinates of the overt leakage area or high-risk coordinates according to the addressing signal to trigger the wall material rupture of the microcapsule at the corresponding location.

[0013] Furthermore, it also includes: A calibration and optimization module is used to statistically analyze the accuracy of high-risk defect determination and the long-term error of the time-series prediction model. If the accuracy rate is less than the accuracy rate threshold, the preset defect strength or the preset attenuation rate is increased. If the long-term error exceeds a preset error tolerance, it is determined that the parameters of the time-series prediction model will be retrained. The retrained parameters include the trainable weights and bias parameters of the long short-term memory network.

[0014] Furthermore, the multi-frequency joint scanning module includes: A pressure control unit, which generates the multi-band pressure excitation; A distributed capacitive sensing unit is used to detect the deformation response signal generated by the thin film region to which the multi-band pressure excitation is applied, and to detect the pressure excitation signal generated according to the multi-band pressure excitation. A signal synchronization unit is used to synchronize and align the deformation response signal with the pressure excitation signal and fuse the data to generate the space leakage probability map.

[0015] Furthermore, the microcapsule system is pre-embedded in the lower support structure or encapsulation gel of the chip film, which are respectively a timely repair capsule system and a high-risk prevention capsule system; The microcapsules contain a single-component or two-component curable repair material.

[0016] Compared with existing technologies, the thin-film airtightness detection device for microfluidic chips of the present invention has the following advantages: It generates a spatial leakage probability map through multi-frequency joint scanning for comprehensive initial screening, which can efficiently distinguish between overt leaks and hidden defects; by applying fractal pressure waves to suspicious areas and analyzing signal attenuation characteristics, it can dynamically adjust detection sensitivity to achieve highly specific capture of early micro / nano-level defects; by constructing a time-series prediction model to quantify the long-term evolution risk of defects, it can intelligently decide whether to perform preventative repairs based on the predicted risk value; and finally, by activating the pre-embedded microcapsule system, it achieves timely repair of overt leaks and early intervention for high-risk hidden defects. This setup constructs a fully automated closed loop from comprehensive screening, precise characterization, risk prediction, intelligent decision-making, to targeted repair, upgrading airtightness maintenance from passive response to proactive prediction and prevention, significantly improving the reliability and lifespan of the chip.

[0017] Furthermore, this invention achieves a synergistic improvement in detection sensitivity and specificity by employing a two-stage excitation strategy that combines multi-frequency joint scanning with fractal excitation. Multi-frequency scanning rapidly locates suspicious areas, while fractal waves, due to their self-similar spectral characteristics, can excite multiple intrinsic response modes of defects over a wide frequency band. Combined with the analysis of signal attenuation gradients, this enables the system to have a highly efficient detection capability for micron or nanometer-scale defects in their nascent stage, effectively overcoming the drawback of traditional single-frequency excitation being prone to missed detections.

[0018] Furthermore, this invention also achieves dynamic optimization of detection sensitivity by introducing an adaptive mechanism based on a response intensity threshold. When the signal attenuation rate is large, it indicates that the material damping characteristics are significant. The system automatically lowers the response intensity threshold, thereby effectively capturing feature signals from deep or weak defects without introducing excessive noise. This avoids missed detections that may occur due to a fixed threshold setting, and enhances the system's adaptability and detection robustness to different materials and chips using different processes.

[0019] Furthermore, this invention achieves a shift from static presence / absence determination to dynamic evolution prediction by constructing a time-series prediction model. This model takes historical sequences of defect response intensity, excitation frequency, and material decay characteristics as input, utilizes a long short-term memory network to mine the temporal correlation between these sequences and defect failure, and outputs a quantified risk value. This enables the device not only to identify current defects in the chip but also to predict its future failure probability and time, providing crucial decision-making support for predictive maintenance.

[0020] Furthermore, this invention achieves scientific and precise repair decisions by setting an intervention risk threshold and linking it to a safety margin. This threshold is not a fixed value, but rather determined based on accelerated aging test data of known defective samples, reserving a specific percentage safety margin before the failure point. A preventative repair instruction is triggered only when the predicted risk value exceeds this dynamic threshold, thereby achieving an optimal balance between over-repair and under-repair, maximizing the cost-effectiveness of repair measures.

[0021] Furthermore, this invention integrates a calibration and optimization module, enabling continuous self-evaluation and iterative evolution of the device. This module automatically triggers two types of optimizations by monitoring the accuracy of high-risk defect detection and the long-term error of the prediction model: first, adjusting the detection threshold to optimize the classification boundary; and second, triggering model parameter retraining to adapt to changes in data distribution. This closed-loop optimization mechanism ensures the stability of the system's performance and its continuously improving adaptability during long-term use.

[0022] Furthermore, this invention achieves precise, minimally invasive, and automated repair by employing a repair strategy that combines a pre-embedded microcapsule system with coordinate-based excitation. The microcapsules are categorized into timely repair and preventative types, matched to different defect risk levels. Repair commands can be directly converted into focused energy excitation at specific coordinates, precisely triggering the release of repair material from the target capsule. This eliminates the need for overall chip coating or disassembly, resulting in a rapid, localized repair process with minimal impact on the overall chip performance. Attached Figure Description

[0023] Figure 1 This is a block diagram of the thin-film airtightness detection device for microfluidic chips in an embodiment of the present invention; Figure 2 This is a schematic flowchart of the thin-film airtightness detection method for microfluidic chips in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the process of determining the corresponding region based on the comparison results of signal strength and signal strength threshold, and the comparison results of noise level and noise level threshold in an embodiment of the present invention. Figure 4This is a flowchart illustrating the process of determining whether to adjust the response intensity threshold based on a comparison between the attenuation rate and a preset attenuation rate in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0025] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0026] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0027] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0028] Please see Figure 1 As shown, it is a module block diagram of the thin film airtightness detection device of the microfluidic chip in an embodiment of the present invention.

[0029] This embodiment includes a multi-frequency joint scanning module, a primary judgment module, a fractal excitation module, a secondary judgment module, a timing decision module, and a repair execution module.

[0030] The multi-frequency joint scanning module includes a pressure control unit, a distributed capacitance sensing unit, and a signal synchronization unit. The pressure control unit applies multi-frequency pressure excitation to the thin film region of the microfluidic chip; the distributed capacitance sensing unit detects the deformation response signal generated by the thin film region under multi-frequency pressure excitation, as well as the pressure excitation signal generated based on the multi-frequency pressure excitation; and the signal synchronization unit synchronizes and fuses the deformation response signal and the pressure excitation signal to generate a space leakage probability map.

[0031] The primary judgment module is connected to the multi-frequency joint scanning module, which is used to perform region-by-region analysis on the spatial leakage probability map. When the signal strength of a certain region is greater than or equal to the signal strength threshold, it is judged as a visible leakage region and a first repair instruction containing the coordinates of the visible leakage region is generated. Timely repair is triggered based on the first repair instruction. When the signal strength of a certain region is less than the signal strength threshold and the noise level is greater than or equal to the noise level threshold, it is judged as a defect-hidden region and the coordinates and signal attenuation characteristics of the region are output.

[0032] The fractal excitation module is connected to the primary judgment module. It is used to apply fractal pressure waves with specific parameters to the defect concealment area and perform dynamic analysis based on signal attenuation characteristics to determine the corresponding attenuation rate. When the attenuation rate is greater than the preset attenuation rate, the system automatically lowers the response intensity threshold used to extract the defect feature signal and calculates and outputs the defect response intensity based on the lowered threshold. When the attenuation rate is less than or equal to the preset attenuation rate, the system determines that it is not necessary to lower the response intensity threshold used to extract the defect feature signal and directly outputs the currently detected defect response intensity.

[0033] The secondary judgment module is connected to the fractal excitation module. It is used to receive the defect response intensity. When the defect response intensity is greater than or equal to the preset defect intensity, it is determined that there is a high-risk hidden defect in the area.

[0034] The timing decision module is connected to the secondary judgment module. It is used to construct a timing prediction model for quantifying the long-term behavior of defects based on the defect response intensity and a pre-stored frequency-current mapping relationship database when a high-risk hidden defect is determined to exist. When the predicted risk value output by the timing prediction model is greater than or equal to the intervention risk threshold, a second repair instruction containing the precise coordinates of the high-risk defect is generated.

[0035] The repair execution module is connected to the primary judgment module and the timing decision module respectively. It is used to locate the corresponding defect coordinates, including the coordinates of the explicit leakage area and the coordinates of the high-risk defect, according to the first repair instruction from the primary judgment module or the second repair instruction from the timing decision module; and to activate the microcapsule system pre-embedded in the corresponding coordinates of the chip to implement local material recombination repair.

[0036] Please see Figure 2 The diagram shown is a flowchart illustrating the thin-film hermeticity detection method for microfluidic chips in an embodiment of the present invention. The process in this embodiment includes at least the following steps: S1: Apply multi-band pressure excitation to the thin film region of the microfluidic chip and simultaneously collect deformation response signals through a capacitive sensing array to generate a spatial leakage probability map. Specifically, a series of dynamic pressure excitations at specific frequencies are applied to the microfluidic chip by the pressure control unit within the multi-frequency joint scanning module. Simultaneously, a distributed capacitive sensing unit detects the nanoscale micro-deformation of the thin film caused by the pressure excitation and its phase response, generating frequency response characteristics for each sensing unit at different frequencies corresponding to its spatial location. This extracts multi-dimensional frequency response characteristics (such as resonant frequency, phase lag angle, and multi-frequency coherence), further constructing the original multi-frequency monitoring data for that point. A signal synchronization unit synchronizes and aligns the deformation response signal with the pressure excitation signal, thereby fusing it with the multi-frequency detection data. Simultaneously, a noise level N based on the background signal is introduced as a confidence weight to suppress false alarms in high-noise areas. All noise-calibrated S values ​​are mapped to physical coordinates and, through interpolation and rendering, stored as a grid-like spatial leakage probability map. This map not only reveals the spatial distribution of leakage risk using color gradients, but its generation logic also ensures that the judgment results for high-noise areas are automatically weakened, achieving a unification of defect location and measurement confidence.

[0037] Furthermore, within the working cycle of the multi-frequency joint scanning module, the system pre-defines a dedicated background measurement window. This window typically lies within the quiet period between two active dynamic pressure excitations or in a stable state where no external target excitation is applied. Within this window, the distributed capacitance sensing unit continuously acquires signals from each sensing channel at the same sampling rate. Within the background measurement window, for each independent spatial sensing point, a background voltage signal sequence {Vi(t)} is acquired over a period of time. The fluctuations in this sequence mainly originate from non-target leakage factors such as environmental vibration, circuit thermal noise, and electromagnetic interference. The standard deviation of this sequence is calculated as a noise intensity statistic, i.e., a direct measure of the noise level at that point. To ensure comparability of noise levels between different sensing units or chips and to facilitate fusion calculations using these as weights, the original noise level is typically normalized to obtain a unitless relative noise level coefficient Ni. Ultimately, each sensing unit's corresponding spatial coordinate point will obtain a noise level N value proportional to its background signal fluctuation. The higher the N value, the greater the background noise at that measurement point, and the lower the inherent reliability of its subsequent monitoring data.

[0038] The noise level N generated through the above process is not a preset constant, but a spatial distribution parameter derived from actual measurements and dynamic calculations. When constructing the spatial leakage probability map, this value is used as a confidence weight, thereby systematically suppressing the interference of outliers in high-noise areas on the overall risk assessment and achieving intelligent detection with adaptive signal-to-noise ratio.

[0039] S2: Identify areas in the spatial leakage probability map where the signal strength is greater than or equal to the signal strength threshold, determine them as explicit leakage areas and trigger timely repair; identify areas where the signal strength is less than the signal strength threshold and the noise level is greater than or equal to the noise level threshold, determine them as hidden defect areas and output the coordinates and signal attenuation characteristics of the hidden defect areas. Please see Figure 3 As shown, it is a flowchart for determining the corresponding region based on the comparison results of signal strength and signal strength threshold and the comparison results of noise level and noise level threshold in an embodiment of the present invention.

[0040] Specifically, the signal strength W and noise level N in the spatial leakage probability map are identified in the primary judgment module, and a signal strength threshold W0 and a noise level threshold N0 are preset. The signal strength W is compared with the signal strength threshold W0 and the noise level N is compared with the noise level threshold N0, and the next step is determined based on the comparison results.

[0041] Furthermore, based on the raw voltage signals acquired by the distributed capacitive sensing units, the voltage output sequence of each sensing unit is acquired during the duration of dynamic pressure excitation at a specific frequency applied by the pressure control unit; this window is called the excitation response window. After filtering out the DC bias, the characteristic values ​​representing the signal oscillation intensity are calculated from the voltage sequence within the excitation response window. The most common and stable method is to calculate the root mean square value of this sequence. After completing a series of scans at different frequencies, each sensing unit will obtain a set of signal intensities S at different frequencies. These can be further fused (e.g., by taking the maximum value or weighted sum at each frequency) to form a comprehensive signal intensity S.

[0042] Under static or quasi-static operating conditions where the pressure control unit does not apply any dynamic excitation, a continuous background noise acquisition window is established. Within this window, the raw voltage output sequence of each distributed capacitive sensing unit is acquired, representing the inherent electrical and environmental fluctuations of the system. After filtering out any potentially extremely low-frequency trend terms from this voltage sequence, its characteristic value representing the intensity of random fluctuations is calculated. The most common and stable method is to calculate the standard deviation of this sequence. After completing the data acquisition and calculation for all sensing units, each unit will obtain an independent noise level N. To establish a unified system-level discrimination benchmark, these levels can be further fused (e.g., by taking the 95th percentile or maximum value of all unit noise levels) to form a comprehensive system noise level N.

[0043] In this embodiment, the signal strength threshold W0 is determined based on the background response data of a standard microfluidic chip in a leak-free state. Specifically, under the premise of stable chip design, thin film materials, and manufacturing processes, standard chip samples that have been rigorously tested and confirmed to be free of any leakage defects are collected. These samples are then scanned using the multi-frequency joint scanning module and the capacitive sensing array to obtain a spatial leakage probability map under standard test conditions. The signal strength values ​​W of all grid cells in the map are extracted to form a standard background sample set. By calculating the upper limit or a specific high quantile (such as the 99.9th percentile) of this sample set, a safety margin is added to the obtained statistical value, and then normalized to eliminate dimensions, establishing it as the signal strength threshold W0. For example, this safety margin can be set by increasing the statistical upper limit benchmark value by 20%, and finally, the normalized signal strength threshold W0 is set to 5.2.

[0044] Furthermore, the noise level threshold N0 is determined based on the background response data of a standard microfluidic chip in a leak-free state. Its core purpose is to define the statistical boundary between the inherent background noise of the system and potential weak anomalous signals. Specifically, it is determined as follows: based on the same standard background sample set used to determine the signal strength threshold W0 (i.e., all grid signal strength values ​​W from defect-free chips), the following statistical method is used: a specific higher percentile (e.g., the 95th or 98th percentile) of this sample set is calculated. This percentile value represents the upper limit of fluctuation in signal strength W at the vast majority of normal locations (e.g., 95%) on a good chip. This fluctuation is mainly caused by inherent factors such as system electronic noise, environmental micro-disturbances, and the microscopic uniformity of the thin film material. This statistical value is established as the benchmark for the noise level threshold N0. To provide a certain degree of discrimination stability, a small safety margin can be added (or the statistical value can be used directly). This safety margin aims to buffer the inherent differences in chip characteristics between batches and slight fluctuations in environmental conditions, thereby establishing a more robust discrimination threshold. The obtained statistical value is appended with a safety margin and then normalized to eliminate dimensions, establishing it as the noise level threshold N0. For example, the safety margin can be set to 5% of the statistical benchmark value. Finally, the normalized noise level threshold N0 is set to 3.5. The process of comparing the signal strength W with the signal strength threshold W0 and the noise level N with the noise level threshold N0 is as follows: If W is greater than or equal to W0, it indicates that the signal response detected at that location statistically far exceeds the normal fluctuation limit of the background response of a defect-free chip. This means the corresponding physical location has a macroscopic or submacroscopic defect in the thin-film structure that could lead to dielectric leakage. This defect, under dynamic pressure excitation, produces a deterministic deformation noise far exceeding the noise level. Therefore, the corresponding location is determined to be a visible leak area, and the visible leak processing unit marks this area, extracts its geometric center coordinates as output coordinates, and generates a first repair command for timely repair. This determination has absolute priority and is independent of the comparison result between the noise level N and the noise level threshold N0 (regardless of whether N is less than, equal to, or greater than N0), and does not change or affect the final determination based on W≥W0. Once the condition is met, the system will automatically mark the area, extract its geometric center coordinates, and generate a repair command to ensure a rapid and clear response to the deterministic physical defect.

[0045] In the chip manufacturing stage, micron-sized capsules containing repair monomers are pre-embedded in the thin-film bonding interface. Once the primary judgment module identifies a visible leak area based on the spatial leakage probability map and outputs precise grid coordinates, it generates a first repair command. Upon receiving the first repair command, the repair execution module locates the coordinates and applies focused ultrasound or a laser of a specific wavelength to directionally stimulate the microcapsules, causing them to rupture and release the repair agent. This agent then seeps into the leak gap through capillary action, rapidly solidifying to achieve precise defect repair.

[0046] If W is less than W0 and N is greater than or equal to N0, it indicates that the signal response detected at this location has statistically significantly exceeded the inherent noise fluctuation range of the system and the intact chip. However, its intensity has not yet reached the confidence level to be definitively determined as a macroscopic leakage. This signal characteristic reflects that the thin film structure at this location may have microscopic mechanical property weakening, submicron-level initial damage, or deep hidden defects. Although these defects have not caused strong overall deformation under the current excitation and detection conditions, they have caused detectable abnormal changes in its dynamic mechanical response (such as stiffness, damping, etc.). Therefore, the corresponding location is determined to be a defect-hidden area, and the area is marked by the hidden defect screening unit to extract the geometric center coordinates as the output coordinates. Furthermore, the original full-band response data of this area recorded in the multi-frequency joint scan is retrieved, and the attenuation curve of the signal amplitude with the excitation frequency is analyzed to calculate the attenuation gradient of the signal in the defect-hidden area in multiple frequency bands as the signal attenuation characteristic.

[0047] If W is less than W0 and N is less than N0, it indicates that the signal response detected at this location is statistically indistinguishable from the background response of a defect-free standard chip. Its fluctuations fall entirely within the normal random distribution range determined by system noise, environmental disturbances, and the microscopic uniformity of the material itself. Therefore, the corresponding location is determined to be a normal region. The airtightness defects present in this normal region are acceptable and will not substantially affect the microfluidic chip during use; therefore, there is no need to activate the microcapsule system embedded at the corresponding coordinates.

[0048] S3: Apply a fractal pressure wave to the defect-hidden area and dynamically extract the signal attenuation characteristics to determine the corresponding attenuation rate. If the attenuation rate is greater than the preset attenuation rate, determine the threshold for reducing the response intensity used to extract the defect feature signal and output the defect response intensity based on the reduced threshold. Specifically, the coordinates of the hidden defect area and the signal attenuation characteristics are input into the waveform generation unit in the fractal excitation module, and the optimized parameter set of the fractal pressure wave is calculated by an adaptive algorithm. The optimized parameter set includes the fractal dimension D, the fundamental frequency f0, the frequency multiplication factor γ, the number of harmonic terms M, and the maximum amplitude Amax.

[0049] According to the geometric series formula fn=f0×γ^n, let n=0,1,2,...,M-1, and successively calculate M discrete frequency components {f0,f1,...,f(M-1)} to form a multi-frequency point set on the spectrum. Based on the power-law attenuation formula An=Amax×fn^(-(2-D)), assign a corresponding amplitude to each frequency component to ensure that the power spectral density of the generated composite wave in the frequency domain satisfies the power-law relationship S(fn)∝fn^{-β}, thus mathematically guaranteeing the self-similarity of the spectrum at different scales, where the spectral exponent β=5-2D.

[0050] Furthermore, the waveform generation unit performs phase randomization processing, adding a random initial phase φu within the range of 0 to 2π to each frequency component to avoid the simultaneous superposition of all sinusoidal peaks, which would generate excessive instantaneous pressure and damage the thin film. In the time-domain synthesis stage, all frequency components are superimposed in the time domain to form a composite pressure wave. Finally, post-processing including windowing, bandpass filtering, and amplitude normalization is performed to generate a fractal pressure wave with self-similar spectral characteristics.

[0051] A fractal pressure wave is applied to the defect-hidden area by a waveform generation unit. The acquired current or vibration signal is preprocessed using a capacitive sensing array in the response analysis unit, performing noise reduction, baseline calibration, and time-domain alignment. A short-time Fourier transform is then used to convert the one-dimensional time-domain signal into a two-dimensional time-frequency distribution map. Based on the Hilbert transform, the signal amplitude envelope curve is extracted from the time-frequency components of a specific or full frequency band. The slope of the exponential decay fitting of the envelope curve within a selected time window is calculated to determine the attenuation rate R.

[0052] Please see Figure 4 As shown, it is a flowchart of the process for determining whether to adjust the response intensity threshold based on the comparison result between the attenuation rate and the preset attenuation rate in an embodiment of the present invention.

[0053] The response analysis unit outputs the calculated attenuation rate R to the threshold adaptation unit, and compares it with a preset attenuation rate R0. Based on the comparison result, it determines whether to reduce the response intensity threshold θ.

[0054] In this embodiment, the preset attenuation rate R0 is determined based on the inherent damping characteristics of a standard microfluidic chip in a leak-free state. Specifically, under the premise of stable chip design, thin film materials, and manufacturing processes, standard chip samples that have been rigorously tested and confirmed to be free of any leakage defects are collected. Standard parameter fractal pressure wave excitation is applied to typical locations in the thin film region of the sample chip through the fractal excitation module. Simultaneously, vibration response signals are acquired and time-frequency analyzed by the response analysis unit. The exponential attenuation slope of the signal envelope at each test location is calculated, and the attenuation rate values ​​at all test locations are extracted to form a standard attenuation rate sample set. By calculating the upper limit or a specific high quantile (e.g., the 99.9th percentile) of this sample set, a certain safety margin is added to the obtained statistical value, and this is established as the preset attenuation rate R0. For example, the safety margin here can be set to 10% of the statistical value. Finally, R0 is set to 1.2 dB / ms. The process of comparing the attenuation rate R with the preset attenuation rate R0 is as follows: If R is greater than R0, it indicates that the signal energy attenuation rate in this region is abnormal and there is a highly damped microscopic defect structure. The detection threshold needs to be reduced to capture the weak defect features masked by attenuation. Therefore, the response intensity threshold θ used to extract the defect feature signal is determined to be reduced.

[0055] At this point, the response intensity threshold θ is reduced by the threshold adaptive unit, and the attenuation rate increment threshold ΔR and the corresponding threshold reduction compensation Δθ are set.

[0056] When the attenuation rate R exceeds the preset attenuation rate R0 by an integer multiple of ΔR, the response intensity threshold θ is reduced stepwise according to a preset fixed step size Δθ, i.e., θ = θ0 - n × Δθ, where n is the largest integer satisfying R ≥ R0 + n × ΔR, and θ0 is the initial response intensity threshold that is normalized to eliminate dimensions.

[0057] At the same time, a minimum response intensity threshold θmin is set as a safety boundary. When the calculated threshold is lower than θmin, then θmin is taken as the actual threshold used.

[0058] For example, ΔR = 0.3dB / ms, Δθ = 0.25 (same unit as θ0), and θmin = 0.5 can be set. When R = 2.1dB / ms and R0 = 1.2dB / ms, n = floor((2.1-1.2) / 0.3) = floor(3.0) = 3 is calculated. Then, the adjusted response intensity threshold θ = θ0 - 3 × 0.25.

[0059] If R is less than or equal to R0, it indicates that the dynamic damping characteristics of the region are consistent with those of an intact film and there are no active hidden defects. Therefore, the response intensity threshold θ used to extract defect feature signals is determined not to be reduced, so as to avoid unnecessary intervention in non-defect factors such as microscopic inhomogeneity of the material.

[0060] Furthermore, the current defect response intensity is output based on the reduced threshold, or the current defect response intensity is output directly without reducing the threshold.

[0061] S4: Identify areas where the defect response intensity is greater than or equal to the preset defect intensity, and determine that there are high-risk hidden defects; Specifically, the defect response intensity C corresponding to each defect concealment area is obtained through the secondary judgment module, and a preset defect intensity C0 is set in advance for comparison. Based on the comparison result, it is determined that there are high-risk concealed defects in areas that are greater than or equal to C0.

[0062] In this embodiment, the preset defect intensity C0 is determined based on the failure risk assessment and historical defect data of the microfluidic chip. Specifically, under the premise of stable chip design, thin film materials, and manufacturing processes, chip samples confirmed by long-term service testing or accelerated aging tests to have functional failures due to hidden defects are collected. Through the fractal excitation module and response analysis unit, the coordinates of the confirmed failure areas in these samples are directionally detected to obtain the equivalent defect response intensity C before failure, forming a high-risk defect sample set. By calculating the lower limit or a specific low quantile (such as the 5th or 10th percentile) of this sample set, this statistical value is used as the intensity benchmark for the high-risk defect response. To ensure the conservatism and reliability of the judgment and prevent defects that are close to the failure threshold from being mistakenly judged as safe, a negative safety margin (i.e., a certain margin is subtracted from the benchmark value) is introduced on this benchmark value and established as the preset defect intensity C0. For example, the safety margin here can be set to 15% of the benchmark value, and finally, for example, C0 = 1.5. The process of comparing the defect response intensity C with the preset defect intensity C0 is as follows: If C is greater than or equal to C0, it indicates that the defect response intensity in this area has reached or exceeded the preset defect intensity determined by the statistical analysis of historical failure data. In subsequent use, there is a high probability that it will develop into an explicit leak or cause structural failure. Therefore, it is determined that there is a high-risk hidden defect in this area.

[0063] If C is less than C0, it indicates that although the abnormal response signal in this area can be detected, its defect response intensity is below the preset defect intensity. Such defects are mostly more microscopic damage or in the early stage of evolution. Therefore, this area is determined to be a low-risk area and no repair instruction is generated.

[0064] S5: When a high-risk hidden defect is determined to exist, a time-series prediction model for quantifying the long-term behavior of the defect is constructed based on the defect response intensity. When the predicted risk value output by the time-series prediction model is greater than or equal to the intervention risk threshold, a second repair instruction containing the coordinates of the high-risk defect is generated by combining the pre-stored frequency-current mapping relationship. In this embodiment, when constructing the time series prediction model, a large number of defect hidden regions that have completed the final state determination are used as the training set, and a long short-term memory network is used for training. The trained model is the time series prediction model, and the risk value can be quantified.

[0065] The specific steps involved in constructing a time series prediction model are as follows: (1) Data acquisition and sequence construction, including: constructing a multivariate time series database as the training basis for the model. This database consists of complete historical data records of the hidden defect areas that have been accumulated in previous detections and whose final state (failure or stability) is clear; Each data record is a time series sample, and its input variables are organized by time step, including: dynamic response data: the sequence of defect response intensity C changing with time; excitation condition data: the sequence of the excitation main frequency fp applied at each detection; material state characteristics: the sequence of attenuation rate R and multi-band attenuation gradient extracted at each detection. The output variables for each data record are the final risk label and key evolution parameters of the defect area, including: risk category label: a binary label (e.g., high-risk hidden area or low-risk area); key temporal features: such as the average slope of the defect response intensity during the observation period, the kurtosis of the sequence, etc., used to quantify the evolutionary behavior of the defect; The collected raw time series data are preprocessed, including: sequence alignment and interpolation: sequences with different detection periods and non-uniform sampling time intervals are aligned to a unified time grid using interpolation methods, and missing values ​​are processed; noise filtering: moving average or low-pass filtering is applied to sequences such as defect response intensity to suppress measurement noise and retain trend information; sequence normalization: time series of each feature dimension are normalized separately (e.g., Z-score normalization based on mean and variance) to eliminate the influence of dimensions.

[0066] (2) Feature engineering and model architecture design: Before inputting the time series data into the model, targeted time series feature engineering is performed, including: Construct time-series lag features: Generate lag values ​​(C(t-1), C(t-2), ..., C(tG)) for each original feature over the past G time steps, so that the model can show that it utilizes historical information. The time step can be set to 30 days for example. Extracting sequence statistical features: Calculate the statistics (e.g., mean, standard deviation, slope) of each feature sequence within the sliding window as supplementary static features; Model selection and architecture design: Given the need to model long-term dependencies between time steps to predict defect evolution trends, a Long Short-Term Memory (LSTM) network is preferred as the core architecture. The specific design includes: Input layer: receiving processed multivariate time series (dimensions are [time step size, number of features]). Long Short-Term Memory Network Stacked Layers: Two to three layers of long short-term memory network units are stacked, each containing a certain number (e.g., 64 or 128) of hidden units, and a Dropout layer is introduced to prevent overfitting, in order to extract and abstract temporal patterns layer by layer. Output layer: Based on the task settings, a fully connected layer is used to map the output of the last time step of the Long Short-Term Memory network to the prediction target (such as risk probability or response strength value at several future time steps).

[0067] (3) Model training and hyperparameter optimization: The complete time series dataset after preprocessing and feature engineering is divided into training set, validation set and test set according to the defect hiding area to ensure the independence of the sequences; Training process: The Long Short-Term Memory (LSTM) network model is trained using the training set data. The network parameters are iteratively updated using the Backpropagation Time-Step Expanded (BPTT) algorithm, with the goal of minimizing the cross-entropy loss between the predicted risk and the true risk label (or the mean squared error between the predicted sequence and the true sequence). Hyperparameter optimization: Key hyperparameters of the model are tuned using strategies such as Bayesian optimization or random search. Core optimization hyperparameters include: the number of layers in the Long Short-Term Memory network, the number of hidden units per layer, the Dropout rate, the learning rate, the training batch size, and the time step length of the input sequence. The training process is monitored using a validation set, overfitting is prevented through early stopping, and the optimal combination of model parameters is selected based on the performance on the validation set.

[0068] Preliminary validation and benchmark establishment: Evaluate the performance of the optimal model on the validation set, and record the benchmark values ​​of its high-risk defect identification accuracy and the long-term prediction error of key indicators as the initial reference for subsequent calibration and optimization.

[0069] After model validation and deployment, and model training and optimization, a comprehensive evaluation is conducted using an independent test set. Evaluation metrics include classification performance metrics and regression performance metrics. Model prediction results (e.g., risk values) need to be compared with the actual long-term behavior of defects (e.g., whether they develop into overt leaks in subsequent monitoring). The model is considered successfully trained when its prediction accuracy, stability, and generalization ability all meet the preset engineering application requirements (e.g., prediction accuracy of high-risk defects > 90%, or the coefficient of determination R² for trend prediction > 0.8). Classification performance metrics include the accuracy, precision, and recall for identifying high-risk defects; regression performance metrics include the long-term error between the predicted sequence and the true sequence, such as root mean square error, mean absolute error, and coefficient of determination (R²).

[0070] Model deployment and closed-loop calibration optimization: Model Deployment and Performance Monitoring: The time-series prediction model, which has passed the admission assessment, is deployed to the online detection system to predict the risk of newly added hidden defect areas in real time. The system simultaneously initiates performance monitoring, continuously collecting long-term observation data of new defect samples to form a comparison database between model predictions and actual evolution results.

[0071] The detection device in this embodiment also includes a calibration and optimization module connected to the secondary judgment module and the time-series decision module respectively. The calibration and optimization module is used to perform the following process: Based on monitoring data, the calibration and optimization logic is executed periodically, including: statistics and comparison: periodically statistically analyzing the actual high-risk defect judgment accuracy and actual long-term prediction error of the model in application; threshold dynamic calibration: if the statistically obtained judgment accuracy is lower than the preset accuracy threshold, it is determined that the current risk identification standard may be too lenient; the system automatically determines and executes the adjustment strategy: increasing the preset defect intensity or increasing the preset decay rate to improve the rigor of high-risk judgment. Model retraining trigger: if the statistically obtained long-term error is greater than the preset error tolerance, it is determined that the model may degrade due to changes in data distribution; the system automatically determines and triggers the parameter retraining of the time-series prediction model. The retraining will be based on a historical database containing new data, and steps (1) to (3) will be re-executed to update the trainable weights and biases of the long short-term memory network and other model parameters to maintain its predictive ability.

[0072] Closed-loop feedback and continuous iteration: Through the above mechanism, a complete closed loop is constructed from monitoring, prediction, validation, calibration or retraining. This module ensures that the time series prediction model is not a static tool, but an intelligent system that can continuously self-optimize with data accumulation and experience feedback, maintaining high accuracy and high reliability in the long term.

[0073] Specifically, the validated model is solidified into an executable time-series prediction model and deployed in the model building unit of the time-series decision module. This model is used to receive real-time or near-real-time time-series data online and output a predicted risk value K for the long-term behavior of defects through the risk quantification unit. An intervention risk threshold K0 is pre-set for comparison. Based on the comparison between the predicted risk value K and the intervention risk threshold K0, the source coordinates are determined based on the comparison results and the frequency-current mapping relationship.

[0074] In this embodiment, the intervention risk threshold K0 is determined based on the failure risk assessment and historical defect evolution data of the microfluidic chip. Specifically, assuming the chip design, thin film materials, manufacturing process, and timing prediction model are stable, chip samples that have undergone long-term service testing or accelerated aging tests and have ultimately failed due to hidden defects are collected. By retrieving the complete historical detection data of the samples before the failure occurred and applying a preset safety time margin (e.g., taking the 10% time point before the failure time), the defect state parameters (including defect response intensity C, decay rate R, etc.) corresponding to that time point are extracted.

[0075] Subsequently, the model building unit inputs these state parameters into the fully trained time-series prediction model, performs a forward inference, and calculates the predicted risk value corresponding to the failure sample at the "safety boundary point." This process is repeated for all failure samples, resulting in a series of predicted risk values ​​that constitute a critical intervention risk sample set. The lower limit or a specific low quantile (e.g., the 10th percentile) of this sample set is calculated and used as a baseline for risk warning. To ensure the timeliness and reliability of the warning, a small positive safety margin is added to this statistical baseline. This safety margin aims to buffer the uncertainty of model predictions and individual differences between chips, thereby establishing a more conservative warning line and ensuring that the system can trigger intervention in time before the risk truly approaches the critical level. After adding a small positive safety margin to the obtained statistical value, it is established as the intervention risk threshold K0. For example, the safety margin here can be set to 5% of the baseline value, and K0 is ultimately set to 0.75. The process of comparing the predicted risk value K with the intervention risk threshold K0 is as follows: If K is greater than or equal to K0, it indicates that the predicted risk of the current defect has reached or exceeded the critical level determined by the statistics of historical failure cases. It is highly likely that the defect will develop into an explicit leak or cause functional loss in the subsequent evolution. Therefore, it is identified as a high-risk defect that requires immediate intervention, triggering the repair decision process to generate a second repair instruction containing the coordinates of the high-risk defect.

[0076] If K is less than K0, it indicates that the current risk level is lower than the general critical value of historical failure cases. The system determines that it is in a continuous prediction state, does not generate a repair instruction for the time being, and continues to make predictions until K is equal to or equal to K0, at which point the repair decision process is triggered.

[0077] In this embodiment, a frequency-current mapping relationship is constructed by collecting electrical response data of standard defect samples and establishing a quantitative correspondence between current response characteristics and defect physical parameters.

[0078] Constructing a frequency-current mapping relationship specifically includes the following steps: (1) Data acquisition and calibration set construction, including: constructing a multivariate mapping database as the physical basis for model training. This database consists of complete test records of standardized samples that have been accurately characterized in terms of defect types and geometric parameters, either artificially prepared or accumulated from historical testing; Each data record is a static mapping sample, whose input variables are organized according to defect attributes and test conditions, including: physical characteristics of defects: defect type (such as microcracks, interface debonding, micropores, etc.), equivalent size (micrometer level), depth and position coordinates in the thin film; excitation parameters: excitation frequency fp applied during the test (covering a wide frequency band of 1kHz-10MHz), excitation amplitude, waveform type; environmental conditions: ambient temperature and humidity during the test.

[0079] The output variables for each data record are the corresponding electrical response characteristics, including: current response characteristics: the current amplitude (usually converted to decibels) and current phase difference measured at a given frequency; derived characteristics: quality factor, resonant frequency offset, harmonic distortion rate, etc.; auxiliary verification data: synchronously acquired capacitance changes, and verification images of defect morphology obtained by optical microscope or scanning electron microscope.

[0080] The collected raw data is preprocessed, including: parameter normalization: geometric parameters such as defect size and location coordinates are normalized according to chip size; logarithmic transformation of frequency is performed to balance dynamic range; outlier removal: based on the 3σ principle or box plot analysis, abnormal response data caused by measurement failure are removed; data augmentation: by adding Gaussian noise and simulating slight parameter perturbations, the dataset is expanded within a physically reasonable range to enhance model robustness.

[0081] (2) Feature engineering and model architecture design: Before inputting data into the model, feature engineering is performed on the defect-response relationship, including: Constructing interactive features: Calculating the product of defect size and excitation frequency, the projection coefficients of position coordinates and chip resonant mode basis functions, etc., to reveal the encoded physical interactions; Extracting frequency domain features: For frequency sweep test data, extract the derivative features of the current-frequency curve, the width of the resonant peak, and the ratio of the attenuation slopes in the low-frequency and high-frequency bands, etc. Defect type embedding: Discrete defect type labels are converted into continuous vector representations through an embedding layer to capture the semantic relationships between different types of defects in the response.

[0082] Model Selection and Architecture Design: Given the need to establish an accurate mapping from high-dimensional, nonlinear defect physical parameters to complex electrical responses, a deep fully connected neural network or a physical information neural network is preferred as the core architecture. Specific design includes: Input layer: Receives processed feature vectors (dimension is [number of features]), including defect parameters, excitation frequency, etc. Feature fusion layer: Design 2 to 3 fully connected layers, each containing 128 to 256 neurons, using the ReLU activation function, and introduce a batch normalization layer to accelerate training; Core mapping layer: Stack 3 to 4 residual blocks, each containing two fully connected layers and skip connections to learn complex nonlinear mapping relationships; Multi-task output layer: It adopts a branch structure, with one branch outputting the predicted value of the current amplitude and the other branch outputting the predicted value of the current phase, allowing the model to learn two related but different physical quantities at the same time; Uncertainty estimation module: A Dropout layer is added after the output layer or a Bayesian neural network framework is used during training to enable the model to output confidence intervals of the predicted values.

[0083] (3) Model training, optimization and inverse model construction: the complete dataset is divided into training set, validation set and test set according to the samples to ensure that data with different test frequencies of the same sample are not split into different sets.

[0084] Training process: The deep neural network is trained using the training set data. The network parameters are iteratively updated using the backpropagation algorithm, with the objective of minimizing the mean square error combination between the model's predicted current response (amplitude and phase) and the actual measured value. The loss function can be designed as a weighted sum of amplitude and phase errors. Hyperparameter optimization: Bayesian optimization or hyperparameter search algorithms are used to fine-tune the key hyperparameters of the model. Core optimization hyperparameters include: network depth and width, Dropout rate, learning rate and its decay strategy, and batch size. The training process is monitored using a validation set, and an early stopping strategy is employed to prevent overfitting. Inverse mapping model construction (for source localization of decision generation units): Based on the trained forward mapping model, construct the inverse mapping model using any of the following methods: Direct training method: Swap the input and output, take the current response characteristics as input and the defect location coordinates as output, and re-collect data to train a dedicated inverse network; Optimization solution method: solidify the forward model, and in practical applications, for the measured current response, use optimization algorithms such as gradient descent or genetic algorithm to iteratively search so that the forward model outputs the defect location parameters that are closest to the measured values; Adversarial generation: Train a generator network, input current response, and directly generate the most probable defect location distribution map.

[0085] (4) Model validation, deployment and online updates. Model validation: After model training and optimization, a comprehensive evaluation is performed using an independent test set. Evaluation metrics include: Forward accuracy: mean absolute error of predicted current amplitude and phase, coefficient of determination (R²); Inverse positioning accuracy: for the inverse model, evaluate its root mean square error of predicted defect location and positioning success rate (error within tolerance); Physical consistency: check whether the response trend predicted by the model conforms to known physical laws (e.g., frequency increase leads to current amplitude decrease).

[0086] Specifically, when the predicted risk value K is greater than or equal to the intervention risk threshold K0, the source is located by combining the frequency-current mapping relationship to generate a second repair instruction containing the coordinates of high-risk defects.

[0087] In this embodiment, the measured current response spectrum (including amplitude and phase sequence) obtained from the latest detection of the current defect is used as input, based on the fixed frequency-current mapping relationship. The decision generation unit performs an iterative optimization process: using the known approximate area or historical location of the defect as the initial search point of the optimization algorithm; in each iteration, a set of defect parameters (e.g., location coordinates, size, etc.) are assumed, and the corresponding theoretical current response spectrum is calculated using the mapping relationship; the difference between the calculated theoretical response spectrum and the measured current response spectrum (e.g., mean square error) is increased, and optimization methods such as gradient descent or genetic algorithms are used to adjust the assumed defect parameters to minimize the difference; when the loss is lower than a preset threshold or the iteration limit is reached, the current optimal defect location parameters are output as the localization result. If the system has a pre-set trained generative adversarial network model, the decision generation unit directly inputs the measured current response spectrum into the generator. The generator outputs a defect probability heatmap covering the chip thin film area. The decision generation unit then extracts the peak region with the highest probability density in the heatmap and calculates its geometric center coordinates as the source coordinates of the defect. The decision generation unit integrates and assesses the confidence level of the above positioning results, and finally encapsulates the confirmed high-risk defect coordinates into the second repair instruction to drive subsequent precise repair operations.

[0088] S6: Locate the coordinates of the visible leak area according to the first repair instruction and the coordinates of the high-risk defect according to the second repair instruction, and activate the microcapsule system pre-embedded in the corresponding coordinates.

[0089] Specifically, the coordinate positioning unit within the repair execution module receives the explicit leakage area from the first repair instruction and the high-risk coordinates from the second repair instruction, and converts them into the corresponding addressing signals for the microcapsule system.

[0090] In this embodiment, after receiving the coordinates of the visible leak area and the high-risk defect determined upstream, the coordinate positioning unit calculates these abstract positions based on the detection coordinate system in real time through real-time coordinate transformation, according to preset spatial calibration parameters (such as translation offset, rotation matrix, and scaling factor), and precisely maps them to the physical coordinate system of the microcapsule repair platform, transforming them into absolute positions based on the platform's mechanical origin. The unit encapsulates the target position along with corresponding repair action commands (such as repair type, material dosage, and residence time) into a digital addressing command stream with a fixed frame structure (typically including a frame header, address segment, data segment, checksum, and frame tail) based on a predefined communication protocol with the repair platform controller (such as a specific instruction set based on TCP / IP or a custom binary serial protocol). This command stream is ultimately sent to the motion controller of the repair platform in the form of electrical or optical signals through a specific hardware interface (such as EtherCAT, RS-485, or high-speed serial port). After parsing the instructions, the controller drives the actuator (e.g., linear motor and servo driver) to move to the target position through its internal closed-loop control algorithm (e.g., PID control), and synchronously triggers the repair material release device (e.g., micro pump or piezoelectric nozzle), thereby achieving high-precision, fully automatic closed-loop control from digital coordinates to physical repair actions.

[0091] Furthermore, based on the addressing signal, focused ultrasound or laser of a specific wavelength is emitted to the coordinates of the overt leak area and the high-risk coordinates through the energy excitation unit to trigger the wall material rupture of the corresponding microcapsule at the corresponding location. The microcapsule system is pre-embedded in the lower support structure or encapsulation colloid of the chip film, which are respectively a timely repair capsule system and a high-risk prevention capsule system. The microcapsule is encapsulated with a single-component or two-component curable repair material.

[0092] In this embodiment, for high-risk preventative microcapsule systems embedded deep within the chip's encapsulating colloid or supporting structure, focused ultrasound is employed. Utilizing the phased array acoustic focusing principle, ultrasonic energy is non-invasively penetrated through the material and focused at a sub-millimeter focal point, triggering the capsule from within through thermal or mechanical cavitation effects. For timely repair microcapsule systems distributed on the surface or shallow layers of the chip film, specific wavelength lasers are emitted. These wavelengths are precisely matched and can be efficiently absorbed by the capsule wall material or built-in photosensitizer, thereby generating selective photothermal effects within a micrometer-scale light spot, achieving precise point-to-point triggering on the surface. This dual-mode triggering strategy based on defect location and depth ensures that single-component or dual-component repair materials can be reliably released and cured at various potential failure points, achieving full coverage from deep prevention to immediate surface repair.

[0093] In this embodiment, the timely repair capsule system uses a single-component thermosetting epoxy resin curable repair material, while the high-risk prevention capsule system uses a two-component room-temperature curable siloxane curable repair material. The single-component epoxy system uses low-viscosity epoxy resin as the main component to achieve deep penetration, a latent curing agent as an auxiliary material to achieve heat-triggered in-situ curing, and toughening agents to reduce shrinkage stress and coupling agents to enhance interfacial bonding, ultimately forming a dense and tough rigid sealing layer. The two-component siloxane system uses isolated encapsulated hydroxyl-terminated polydimethylsiloxane mixed with a catalyst-containing crosslinking agent, which undergoes a moisture-catalyzed condensation reaction to cure and form a highly elastic silicone rubber sealing layer, effectively compensating for thermal stress.

[0094] Both component systems, after the microcapsules rupture and release, can fully wet and fill micron or nanometer-scale defect channels due to their low viscosity. This allows them to form a solid barrier with strong adhesion and high density to the substrate through a chemical curing reaction, permanently blocking the penetration paths of gases and moisture, thus achieving reliable repair and long-term maintenance of the chip's airtightness. Specifically, the single-component epoxy system comprises: 60%–85% low-viscosity epoxy resin, 3%–10% latent curing agent, 5%–15% toughening agent, and 1%–3% coupling agent by weight. The two-component siloxane system comprises: 85%–95% hydroxyl-terminated polydimethylsiloxane and 5%–15% crosslinking agent containing a catalyst. It should be noted that the composition and proportions of the single-component and two-component repair materials are existing technologies.

[0095] All technologies not mentioned in the above embodiments are existing technologies. It is understood that no specific limitation is made to any preset parameter or critical parameter in the embodiments of the present invention, and the above values ​​are not limited thereto. Those skilled in the art can adjust the preset parameters or critical parameters accordingly based on actual needs, analysis of historical data, or equipment usage.

[0096] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A thin-film hermeticity detection device for a microfluidic chip, characterized in that, include: The multi-frequency joint scanning module is used to apply multi-band pressure excitation to the thin film region of the microfluidic chip and simultaneously acquire deformation response signals through a capacitive sensing array to generate a spatial leakage probability map. The primary determination module is used to identify regions in the spatial leakage probability map where the signal strength is greater than or equal to the signal strength threshold, determine them as explicit leakage regions and generate a first repair instruction containing the coordinates of the explicit leakage regions, and identify regions where the signal strength is less than the signal strength threshold and the noise level is greater than or equal to the noise level threshold, determine them as hidden defect regions and output the coordinates and signal attenuation characteristics of the hidden defect regions. The fractal excitation module is used to apply fractal pressure waves to the defect concealment area and dynamically extract the signal attenuation characteristics to determine the corresponding attenuation rate. If the attenuation rate is greater than the preset attenuation rate, the response intensity threshold for extracting the defect feature signal is reduced and the defect response intensity is output based on the reduced threshold. The secondary judgment module is used to identify areas where the defect response intensity is greater than or equal to the preset defect intensity, and to determine that there is a high-risk hidden defect. The timing decision module is used to construct a timing prediction model that quantifies the long-term behavior of the defect based on the defect response intensity when the existence of the high-risk hidden defect is determined, and to generate a second repair instruction containing the coordinates of the high-risk defect by combining the pre-stored frequency-current mapping relationship when the predicted risk value output by the timing prediction model is greater than or equal to the intervention risk threshold. The repair execution module is used to locate the coordinates of the explicit leakage area according to the first repair instruction and the coordinates of the high-risk defect according to the second repair instruction, and activate the microcapsule system pre-embedded in the corresponding coordinates.

2. The thin-film hermeticity detection device for microfluidic chips according to claim 1, characterized in that, The primary determination module includes: The overt leakage processing unit is used to mark the area where the signal strength is greater than or equal to the signal strength threshold as an overt leakage area, extract the geometric center coordinates as output coordinates, generate the first repair instruction, and activate the pre-embedded microcapsule system for timely repair based on the first repair instruction. The hidden defect screening unit is used to mark regions where the signal strength is less than the signal strength threshold and the noise level is greater than or equal to the noise level threshold as hidden defect regions, extract the geometric center coordinates as output coordinates, and calculate the signal attenuation gradient of the hidden defect region in multiple frequency bands as the signal attenuation feature.

3. The thin-film hermeticity detection device for microfluidic chips according to claim 2, characterized in that, The fractal excitation module includes: A waveform generation unit is used to generate the fractal pressure wave having self-similar spectral characteristics; A response analysis unit is used to perform time-frequency analysis on the current or vibration signal acquired after the fractal pressure wave is applied, in order to determine the attenuation rate of the calculated signal envelope; A threshold adaptive unit is used to reduce the response intensity threshold by a preset ratio when the attenuation rate is greater than the preset attenuation rate.

4. The thin-film hermeticity detection device for microfluidic chips according to claim 3, characterized in that, The fractal excitation module is also used to determine the direct output of the defect response intensity based on the attenuation rate being less than or equal to the preset attenuation rate.

5. The thin-film hermeticity detection device for microfluidic chips according to claim 1, characterized in that, The timing decision module includes: The model building unit is used to construct the time-series prediction model based on the time series of the defect response intensity, the corresponding excitation frequency, and the signal attenuation characteristics as inputs, using a long short-term memory network. The risk quantification unit is used to run the time series prediction model and output the predicted risk value; A decision generation unit is used to perform source localization by combining the frequency-current mapping relationship when the predicted risk value is greater than or equal to the intervention risk threshold, so as to generate a second repair instruction containing the coordinates of the high-risk defect.

6. The thin-film hermeticity detection device for microfluidic chips according to claim 5, characterized in that, The model building unit is used to perform accelerated aging tests on known defect samples and record the corresponding failure times; The model building unit is also used to input the defect state parameters corresponding to a specific percentage safety margin before the failure time into the time-series prediction model, and the output risk value is the intervention risk threshold.

7. The thin-film hermeticity detection device for microfluidic chips according to claim 1, characterized in that, The repair execution module includes: The coordinate positioning unit is used to receive the coordinates of the explicit leakage area in the first repair instruction and the high-risk coordinates in the second repair instruction, and convert them into addressing signals corresponding to the microcapsule system. An energy excitation unit is used to emit focused ultrasound or a laser of a specific wavelength to the coordinates of the overt leakage area or high-risk coordinates according to the addressing signal to trigger the wall material rupture of the microcapsule at the corresponding location.

8. The thin-film hermeticity detection device for microfluidic chips according to claim 1, characterized in that, Also includes: A calibration and optimization module is used to statistically analyze the accuracy of high-risk defect determination and the long-term error of the time-series prediction model. If the accuracy rate is less than the accuracy rate threshold, the preset defect strength or the preset attenuation rate is increased. If the long-term error exceeds a preset error tolerance, it is determined that the parameters of the time-series prediction model will be retrained. The retrained parameters include the trainable weights and bias parameters of the long short-term memory network.

9. The thin-film hermeticity detection device for microfluidic chips according to claim 1, characterized in that, The multi-frequency joint scanning module includes: A pressure control unit, which generates the multi-band pressure excitation; A distributed capacitive sensing unit is used to detect the deformation response signal generated by the thin film region to which the multi-band pressure excitation is applied, and to detect the pressure excitation signal generated according to the multi-band pressure excitation. A signal synchronization unit is used to synchronize and align the deformation response signal with the pressure excitation signal and fuse the data to generate the space leakage probability map.

10. The thin-film hermeticity detection device for microfluidic chips according to claim 1, characterized in that, The microcapsule system is embedded in the lower support structure of the chip film or in the encapsulation gel, and is respectively a timely repair capsule system and a high-risk prevention capsule system; The microcapsules contain a single-component or two-component curable repair material.

Images