A sealing tester and intelligent pressure gradient regulation method thereof

Abstract

The application provides a sealing tester and a pressure gradient intelligent regulation method thereof. The scheme comprises the following steps: collecting and normalizing multi-dimensional time sequence pressure data, extracting dimensionless pressure response characteristics to construct a fingerprint vector, clustering and center calibration based on a historical non-leakage mode, realizing sealing state confidence calculation by combining Mahalanobis distance and local consistency algorithm, triggering a perturbation verification if it is determined to be a critical state, identifying a leakage type through positive pressure pulse excitation and dynamic pressure recovery analysis, and dynamically updating a historical mode library. The method can eliminate the influence of material thickness, environment and sensor offset, realize intelligent high-precision discrimination of the sealing state of the blister plate, and enhance the model adaptability and self-learning ability.

Description

Technical Field

[0001] This invention relates to the field of medical blister testing, and more particularly to a sealing tester and its intelligent pressure gradient control method. Background Technology

[0002] Current non-destructive sealing tests for medical blister packs generally use a method of comparing fixed-point pressure changes during the vacuuming process with a fixed threshold to determine leakage. While this method has advantages such as being non-contact and rapid, the fixed threshold relies on pre-calibration and empirical ranges, making it difficult to adapt to differences in nonlinear pressure response caused by variations in membrane thickness, molding depth, material elastic modulus, and temperature and humidity fluctuations. It is also sensitive to critical micro-leakage and batch fluctuations, easily leading to missed detections and misjudgments. At the same time, existing systems lack the ability to mine patterns in the dynamic evolution of multi-point pressure data, failing to fully utilize the high-dimensional features such as morphology, entropy, and nonlinear coupling inherent in the pressure time series. This makes it difficult to capture the global mechanical response characteristics of the blister pack under negative pressure, resulting in low detection sensitivity, insufficient generalization ability, and the need for repeated manual calibration when switching to new materials and structures, leading to high conversion costs.

[0003] Therefore, a leakage detection scheme is needed that can automatically extract response fingerprints from multi-point pressure time series, dynamically model different material / thickness / structural characteristics, eliminate the need for manually setting fixed judgment thresholds, and adaptively update the pattern library online. Claim 1 of this application solves the problems of high false positive rate, low sensitivity, and difficulty in adaptive generalization of fixed threshold discrimination by generating pressure response fingerprints, constructing a historical pattern library, calculating the confidence level of the sealing state, and triggering perturbation verification at the critical state. Summary of the Invention

[0004] To overcome the high false positive rate and low sensitivity of traditional threshold discrimination techniques, this application provides a sealing tester and its intelligent pressure gradient control method to achieve highly sensitive and robust detection of critical micro-leakage and nonlinear material response.

[0005] This application provides a method for intelligent control of pressure gradients, comprising:

[0006] The pressure data collected by each bubble sensing unit in the flexible pressure sensor array during the vacuuming process is obtained, and the pressure data is normalized and preprocessed to generate relative pressure change data.

[0007] Extract the response features of the relative pressure change data to generate the pressure response fingerprint of each blister unit;

[0008] A historical pattern library is constructed based on historical leak-free benchmark test data. The historical pattern library is divided into several typical response pattern clusters by density peak clustering algorithm, the center position of each cluster is determined, and a set of historical pattern cluster centers with process labels is generated.

[0009] The pressure response fingerprint is compared with the historical pattern library to calculate the sealing confidence of each blister unit.

[0010] Determine whether the confidence level of the sealing state meets the preset judgment conditions;

[0011] If the confidence level of the sealing state meets the preset judgment condition, then the corresponding blister unit is determined to be well sealed;

[0012] If the confidence level of the sealing state does not meet the preset judgment condition, a perturbation verification is triggered and a perturbation verification control signal is generated. In response to the perturbation verification control signal, a positive pressure pulse is injected into the sealing chamber and pressure recovery data is collected. The final leakage judgment result of the corresponding blister unit is determined based on the pressure recovery data.

[0013] This application also provides a sealing tester for performing the above-described intelligent pressure gradient control method, including:

[0014] The sealed enclosure forms a closed space for accommodating the blister pack;

[0015] A vacuum pumping assembly, connected to the sealed enclosure, is used to evacuate the sealed enclosure and adjust the vacuum level inside the sealed enclosure.

[0016] A flexible pressure sensor array is disposed inside the sealed box and above the blister pack, and has multiple sensing units arranged corresponding to each blister pack on the blister pack.

[0017] A lifting mechanism, located inside the sealed box, is used to push the blister packs toward the flexible pressure sensor array, so that the top of each blister pack is pressed into contact with the flexible pressure sensor array.

[0018] A positioning platform is set at the moving end of the lifting mechanism to fix the blister pack and align each blister with the corresponding sensing unit in the flexible pressure sensor array.

[0019] The controller is electrically connected to the lifting mechanism, the flexible pressure sensor array, and the vacuum assembly, respectively, and is used to control their coordinated actions; wherein, the flexible pressure sensor array is configured to detect the pressure changes at each sensing unit before and after the sealed box is vacuumed, so as to reflect the sealing status of each blister point by point.

[0020] Compared with the prior art, the beneficial effects of this application are:

[0021] This solution transforms traditional fixed pressure difference threshold discrimination into adaptive matching based on fingerprints and pattern clusters by constructing a pressure response fingerprint and historical pattern library. This effectively overcomes the problems of high false positive rates and poor generalization caused by differences in materials, thicknesses, and structures of various specifications, achieving zero calibration and cross-batch applicability. Simultaneously, it automatically triggers micro-positive pressure pulse perturbation verification for blister packs in a critical response state. By analyzing the spatiotemporal coupling characteristics of the pressure recovery trajectory, it accurately identifies micro-leakage and weak edge sealing, significantly improving detection sensitivity and reducing false alarm rates. The overall solution possesses self-learning and updating capabilities, real-time feedback, and physical interpretability, balancing high precision, strong robustness, and rapid changeover requirements. It is suitable for large-scale intelligent quality control scenarios in various medical blister pack production lines. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of one or more embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart of the intelligent pressure gradient control method is shown;

[0024] Figure 2 A schematic diagram of the sealing performance tester is shown.

[0025] Figure 3 A schematic diagram of the internal structure of the sealed enclosure is shown;

[0026] Figure 4 A structural schematic diagram of the interior of the sealed enclosure from another angle is shown.

[0027] Explanation of key component symbols:

[0028] 100-Sealed housing; 200-Vacuum assembly; 210-Vacuum pump; 220-Vacuum proportional valve; 300-Flexible pressure sensor array; 400-Lifting mechanism; 410-Servo motor; 420-Transmission assembly; 421-Synchronous belt drive structure; 422-Ball screw; 423-Screw nut; 430-First support plate; 500-Positioning stage; 600-Frame; 700-Second support plate; 800-Guide rod; 900-Buffer spring. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0030] Example 1

[0031] Please see Figure 1 This application provides a method for intelligent control of pressure gradient, specifically including:

[0032] S1: Obtain the high-dimensional time-series pressure data matrix collected by the flexible pressure sensor array for each bubble sensing unit during the entire vacuuming process, and perform normalization preprocessing based on the high-dimensional time-series pressure data matrix to eliminate the influence of absolute pressure offset, and generate a relative pressure change data sequence.

[0033] S2: Perform key morphological feature extraction on the relative pressure change data sequence, calculate multidimensional dimensionless parameters including initial contact slope, maximum response delay time, peak rise rate and plateau period fluctuation entropy, and generate pressure response fingerprint of a single bubble cap unit.

[0034] S3: Construct a historical pattern library under leak-free conditions by using multiple sets of pressure response fingerprints accumulated from historical benchmark tests. Divide the historical pattern library into several typical response pattern clusters using density peak clustering algorithm and determine the center position of each cluster, generating a set of historical pattern cluster centers with process labels.

[0035] S4: Input the pressure response fingerprint generated by the current blister unit under test into the discrimination logic, calculate the Mahalanobis distance and local neighborhood consistency score between the pressure response fingerprint and each cluster center in the historical pattern cluster center set, and generate the sealing state confidence index.

[0036] S5: Determine whether the confidence index of the sealing state meets the first judgment condition of falling into the known leak-free mode cluster and having a confidence level higher than the preset threshold. If the first judgment condition is not met, it is determined to be a critical response state and a perturbation verification command is triggered to generate a perturbation verification control signal.

[0037] S6: Based on the perturbation verification control signal, a small amount of controllable positive pressure pulse is injected into the sealed box and pressure recovery trajectory data is collected simultaneously. The spatiotemporal coupling characteristics of the pressure recovery trajectory data are analyzed to identify local damping anomalies and generate the final leakage judgment result.

[0038] S7: Update the sample distribution information in the historical pattern library and the central set of historical pattern clusters based on the final leakage discrimination results. If outlier fingerprint vectors caused by new materials or structures are detected, they are marked as seed points of new categories to optimize the division boundary of subsequent typical response pattern clusters and generate dynamically updated adaptive discrimination model parameters.

[0039] Step S1: Obtain a high-dimensional time-series pressure data matrix collected by the flexible pressure sensor array for each bubble-shaped sensing unit throughout the vacuuming process. Perform normalization preprocessing based on this high-dimensional time-series pressure data matrix to eliminate the influence of absolute pressure offset, generating a relative pressure change data sequence. Specifically, this includes:

[0040] S1.1: Based on the spatial layout coordinates and sampling time axis of the flexible pressure sensor array, the raw voltage signals output by each bubble sensing unit during the period from the start of vacuuming to the steady state are synchronously acquired and time-series aligned to construct a high-dimensional time-series pressure data matrix containing spatial coordinate information and time-series sampling point sequence.

[0041] Sub-step S1.1 aims to construct a high-dimensional temporal pressure data matrix containing spatial and temporal dimensions, providing standardized input for subsequent feature extraction. The system initializes the sampling clock of the flexible pressure sensor array, setting the sampling frequency to 100Hz to 500Hz to ensure coverage of the high-frequency dynamic response of the blister deformation. A hardware interrupt signal is triggered at the start of vacuuming, synchronously initiating the analog-to-digital conversion process of all sensing units to eliminate time deviations in multi-channel acquisition. The raw voltage signal output by each sensing unit is read in real time, reflecting the instantaneous contact pressure of the blister top against the sensor. The read voltage values ​​are mapped to the physical layout coordinates of the sensor array to establish a spatial index relationship, determining the blister position corresponding to each data point. Based on a preset sampling time axis, the continuously acquired voltage data are arranged in timestamp order to form a one-dimensional time series. Length verification is performed on the sensing unit sequence corresponding to each blister. If data is missing due to communication packet loss, linear interpolation is used to fill in the missing points, ensuring consistency in the temporal dimension. The time series data of all bubble cap units are integrated to construct a three-dimensional data matrix, where the first dimension represents the sensor row coordinates, the second dimension represents the column coordinates, and the third dimension represents the time sampling points. Through the above synchronous acquisition and time-series alignment processing, the discrete voltage signals from the previous step are transformed into a high-dimensional time-series pressure data matrix with a complete spatiotemporal topology, realizing the accurate mapping of detection data from physical signals to digital models and laying the data foundation for eliminating absolute pressure offset.

[0042] For example, for PVC medical blister packs, the sensor array is set to 8 rows and 12 columns, with a total of 96 sensing units. The sampling frequency is set to 200Hz, the vacuuming process lasts for 60 seconds, and the corresponding number of sampling points in the time dimension is 12000. Acquisition is triggered at t=0s, and each sensing unit outputs a 0-5V analog voltage, which is converted into a digital value of 0-4095 by a 12-bit ADC. If the sensor at (3,5) experiences data loss at t=10.5s, linear interpolation is performed using the voltage values ​​at t=10.4s and t=10.6s to complete the data. Finally, a three-dimensional matrix M with a size of 8×12×12000 is generated, where M[i][j][k] represents the pressure value of the sensor in the i-th row and j-th column at the k-th sampling point. This matrix completely records the dynamic deformation process of the blister pack in a vacuum environment, significantly improving the spatiotemporal consistency of the data and effectively supporting subsequent high-precision feature analysis.

[0043] S1.2: Based on the initial contact pressure reference value of each sensing unit in the high-dimensional time-series pressure data matrix, perform zero-point drift correction operation on all time-series data in the matrix to eliminate the absolute pressure offset component caused by the difference in sensor preload, and generate a zero-bias correction time-series pressure data matrix.

[0044] S1.3: Based on the maximum dynamic response amplitude in the time-series pressure data matrix after zero bias correction, perform amplitude normalization mapping on the pressure change curve of each blister sensing unit in the matrix to unify the pressure response dimensions of blister cells with different material thicknesses and generate a standardized relative pressure change data sequence.

[0045] Obtain the time-series pressure data matrix after zero-bias correction, and extract the pressure response curve of each blister sensing unit during the entire vacuuming process. Traverse the time-series data, retrieve the maximum peak and minimum valley values ​​of the pressure for each sensing unit, and calculate the difference between them as the maximum dynamic response amplitude of that unit. Construct a normalized mapping function, subtracting the minimum valley value from the original pressure value and dividing by the maximum dynamic response amplitude, to eliminate the difference in absolute pressure dimensions caused by different material thicknesses. Perform the above mapping operation on all sensing units within the matrix, uniformly compressing the pressure variation range at each point to a dimensionless interval of 0 to 1. Generate a standardized relative pressure change data sequence, preserving the morphological characteristics of the pressure response while eliminating amplitude deviations. Through amplitude normalization mapping processing, the result of the previous step is transformed into a standardized relative pressure change data sequence with uniform dimensions, achieving the expected technical effect of comparability of the response characteristics of blister packs of different specifications.

[0046] For example, for a PVC blister pack with a thickness of 0.25 mm, the sensor collects a raw pressure sequence with a maximum value of 1200 Pa, a minimum value of 200 Pa, and a maximum dynamic response amplitude of 1000 Pa. For a pressure value of 800 Pa at a certain moment in the sequence, substituting it into the normalization formula... The normalized value was calculated to be 0.6. For a blister pack of the same material with a thickness of 0.15 mm, its maximum response amplitude is only 600 Pa. After the same processing, its pressure change pattern is mapped to the same range. This processing eliminates the difference in signal amplitude caused by the difference in film thickness, allowing the subsequent feature extraction algorithm to focus on the morphological features such as the slope and delay of the response curve, significantly improving the algorithm's generalization ability to products of different specifications.

[0047] S1.4: Based on the temporal continuity characteristics of the standardized relative pressure change data sequence, a sliding window filter is performed to smooth the transient noise spikes in the sequence in order to remove abnormal data points caused by environmental vibration interference and generate a denoised and clean relative pressure change data sequence.

[0048] S1.5: Based on the shape preservation characteristics of the clean relative pressure change data sequence after denoising, the sequence data is resampled and interpolated to unify the time resolution, and finally outputs a relative pressure change data sequence that eliminates the influence of absolute pressure offset and has consistent spatiotemporal dimensions.

[0049] Step S2: Extract key morphological features from the relative pressure change data sequence, calculate multidimensional dimensionless parameters including initial contact slope, maximum response delay time, peak rise rate, and plateau period fluctuation entropy, and generate a pressure response fingerprint for a single bubble cap unit. Specifically, this includes:

[0050] S2.1: Perform time-domain differential operation on the relative pressure change data sequence to obtain the initial contact slope parameter reflecting the mechanical characteristics at the moment of contact between the bubble and the sensor, and the maximum response delay time parameter characterizing the hysteresis effect of gas expansion response.

[0051] Receive the denoised and clean relative pressure change data sequence generated by the S1 step preprocessing. This sequence has eliminated the absolute pressure offset and unified the time resolution, and is used as the input object for time-domain differentiation operations.

[0052] S2.2: Based on the dynamic time window defined by the initial contact slope parameter and the maximum response delay time parameter, extreme value search and gradient accumulation calculation are performed on the rising phase data segment in the relative pressure change data sequence to generate the peak rise rate parameter describing the speed of blister bulging.

[0053] Receive the initial contact slope parameter and maximum response delay time parameter generated by step S2.1, as well as the denoised and clean relative pressure change data sequence output by step S1.

[0054] The initial rise time of the pressure response curve is determined based on the maximum response delay time parameter, and the stable range of mechanical transmission is determined by combining the initial contact slope parameter. The time window that includes the main deformation process of blister bulging is dynamically defined.

[0055] Within a defined dynamic time window, a first-order difference operation is performed on the relative pressure change data sequence to calculate the pressure increment between adjacent sampling points and construct a gradient sequence that reflects the instantaneous deformation rate.

[0056] A sliding extremum search algorithm is used to traverse the gradient sequence, identify local maxima as potential peak candidates, and eliminate pseudo-extremes caused by high-frequency noise from the sensor, retaining the real pressure jump points with physical significance.

[0057] The effective gradient values ​​after screening are accumulated and integrated to calculate the total pressure change per unit time, so as to quantify the overall inflating efficiency of the bubble under vacuum negative pressure.

[0058] S2.3: Using the steady-state interval boundary determined by the peak rise rate parameter, the Shannon entropy estimation algorithm is applied to the plateau period data segment in the relative pressure change data sequence to quantify the degree of disorder in pressure fluctuations and generate plateau period fluctuation entropy parameters.

[0059] Obtain the steady-state interval boundary defined by the peak rise rate parameter, and extract the plateau pressure data segment corresponding to the time window from the relative pressure change data sequence.

[0060] Discrete probability distribution modeling is performed on the extracted platform period pressure data segment. The normalized pressure amplitude is divided into several probability bins of equal width, and the frequency of occurrence of sampling points in each probability bin is counted.

[0061] S2.4: Combining the initial contact slope parameter, maximum response delay time parameter, peak rise rate parameter, and plateau period fluctuation entropy parameter, additional morphological characteristic parameters, including the location of the attenuation inflection point, are obtained to construct a multidimensional dimensionless feature set. These additional morphological characteristic parameters include, but are not limited to, the location of the attenuation inflection point, obtained by taking the second derivative of the pressure attenuation stage data, and are used to characterize the material's elastic recovery hysteresis properties.

[0062] Based on the extracted initial contact slope, maximum response delay time, peak rise rate, and plateau period fluctuation entropy parameters, the pressure decay initiation point in the relative pressure change data sequence is located. Second derivative calculations are performed on the data segments of the pressure decay phase to identify curvature extrema and determine the decay inflection point, quantifying the elastic recovery hysteresis characteristics of the blister material after negative pressure release.

[0063] The time difference between the decay inflection point and the peak time is calculated as a recovery delay parameter, reflecting the dynamic resistance to gas exchange inside the sealed cavity. An integral operation is performed on the complete pressure response curve, and the area enclosed by the curve is obtained as the energy dissipation index, characterizing the total viscoelastic loss during the bubble deformation process.

[0064] A sliding window standard deviation algorithm is used to calculate the local fluctuation coefficients at the rising and falling edges, respectively, to construct a morphological symmetry index to evaluate the mechanical uniformity of the blister bulging process. The first local minimum value after the zero-crossing of the pressure sequence and its corresponding timestamp are extracted to generate a residual stress retention rate parameter, which is used to identify the decrease in pressure maintenance capability caused by minor leaks.

[0065] S2.5: Perform vector space mapping operation on the multidimensional dimensionless feature set to encapsulate the scattered feature parameters into a single bubble-cell pressure response fingerprint with a unique identifier, which serves as a standardized input object for subsequent clustering analysis and similarity comparison.

[0066] Step S3: Construct a historical pattern library under leak-free conditions using multiple sets of pressure response fingerprints accumulated from historical benchmark tests. Divide the historical pattern library into several typical response pattern clusters using a density peak clustering algorithm and determine the center position of each cluster, generating a set of historical pattern cluster centers with process labels. Specifically, this includes:

[0067] S3.1: Perform data cleaning and outlier removal on multiple sets of pressure response fingerprints collected and preprocessed during the historical benchmark testing phase to eliminate interference from outlier samples caused by instantaneous sensor noise or environmental fluctuations, and generate a high-confidence standard pressure response fingerprint set.

[0068] S3.2: Calculate local density distribution index and relative distance metric index based on high confidence standard pressure response fingerprint set to quantify the degree of clustering of each sample point in the feature space and the degree of separation from other sample points, and generate a weighted feature data matrix containing density weight information and distance weight information.

[0069] Receive a set of high-confidence standard pressure response fingerprints cleaned by S3.1, which serves as the input data source for the density peak clustering algorithm.

[0070] Construct the Euclidean distance matrix in the multidimensional feature space, calculate the geometric distance between any two fingerprint vectors, and quantify the degree of separation between samples.

[0071] S3.3: Utilize the weighted feature data matrix to perform density peak clustering iterative operation, automatically identify local density maxima as initial clustering seeds, and assign the remaining samples to their respective clusters based on distance gradients, generating several initially divided clusters of typical response patterns without leakage and their corresponding subsets of samples within each cluster.

[0072] Based on the weighted feature data matrix, the local density value of each sample point is calculated by traversing it. The number of samples in the neighborhood is counted according to the cutoff distance parameter, and a local density distribution vector reflecting the degree of sample clustering is generated.

[0073] Perform a descending sort operation on the local density distribution vector to determine the relative ranking of each sample point in the density dimension, and identify the candidate cluster center point with the highest local density.

[0074] Calculate the minimum distance from each sample point to all higher-density sample points, and use this as a relative distance index to quantify the degree of separation between the sample and the high-density region.

[0075] By multiplying the local density value by the relative distance index, a decision graph coordinate system is constructed, and extreme points with both high local density and high relative distance are automatically selected as initial clustering seeds.

[0076] Based on the distance gradient principle, the remaining non-seed sample points are assigned to the clusters of the nearest and densest seed points, forming a preliminary division of the typical response pattern clusters without leakage and their corresponding intra-cluster sample subsets.

[0077] Through density peak clustering iterative operations, weighted feature data is transformed into a set of typical response pattern clusters with clear structure, thereby realizing automatic classification and boundary delineation of blister mechanical response patterns in an unsupervised state.

[0078] For example, the cutoff distance was set to 2% of the average distance in the feature space, and local density was calculated for 1000 fingerprint vectors. Three high-density extrema were identified as seeds, corresponding to PVC thin-walled, PVDC thick-walled, and composite membrane structures, respectively. The remaining 997 samples were assigned to the nearest high-density seed clusters according to the distance gradient, forming three independent pattern clusters. The homogeneity of samples within each cluster was significantly improved, effectively eliminating the classification bias caused by manually set thresholds.

[0079] S3.4: Perform cluster centroid coordinate calculation on a subset of samples within several leakage-free typical response mode clusters to obtain the core vector representing the geometric center and statistical characteristics of each cluster, and generate the initial cluster center position vector for each typical response mode cluster.

[0080] S3.5: Based on the initial cluster center location vector, associate the process batch information, measured film thickness values ​​and environmental temperature and humidity labels in the original benchmark test to give each cluster center a clear physical attribute identifier, and generate a set of historical pattern cluster centers with complete process labels.

[0081] Step S4: Input the pressure response fingerprint generated by the current blister unit under test into the discrimination logic, calculate the Mahalanobis distance and local neighborhood consistency score between the pressure response fingerprint and each cluster center in the historical pattern cluster center set, and generate a sealing state confidence index. Specifically, this includes:

[0082] S4.1: Obtain the pressure response fingerprint generated by the current blister unit under test and the set of historical pattern cluster centers with process labels. Based on the Mahalanobis distance algorithm, perform covariance normalization on the Euclidean distance between the pressure response fingerprint and each cluster center in the historical pattern cluster center set to eliminate the influence of dimensional differences and correlations between different feature dimensions, and generate a multidimensional spatial statistical distance sequence.

[0083] Retrieve a subset of target clusters from the historical pattern cluster center set that matches the current blister process label to be tested, and extract the multidimensional dimensionless feature parameters of all samples in the subset to form a local sample matrix.

[0084] Calculate the mean vector and covariance matrix of each feature dimension in the local sample matrix to construct the basis of a statistical model that characterizes the feature distribution pattern under the leak-free state.

[0085] The covariance matrix is ​​regularized using singular value decomposition, and a small perturbation term is added to ensure matrix invertibility, thereby generating a stable inverse covariance matrix.

[0086] The difference vector between the pressure response fingerprint of the current blister under test and the center of the target cluster is used as input, and a quadratic operation is performed in combination with the inverse covariance matrix to calculate the Mahalanobis distance value, which reflects the degree of deviation of multidimensional features.

[0087] S4.2: Based on the multidimensional spatial statistical distance sequence, the target nearest neighbor cluster center is determined. The local neighborhood consistency score calculation model is used to perform topological structure analysis on the sample point distribution density within the preset radius of the target nearest neighbor cluster center to evaluate the local clustering degree of the current pressure response fingerprint in the feature space and generate a local neighborhood consistency score scalar.

[0088] Based on the multidimensional spatial statistical distance sequence, the target historical pattern cluster center with the minimum Mahalanobis distance value is extracted and defined as the nearest neighbor reference benchmark of the current blister fingerprint vector to be tested.

[0089] Using the nearest neighbor cluster center of the target as the geometric origin, a hypersphere neighborhood search range with a preset threshold R is constructed in the multidimensional feature space. This radius R is dynamically set based on the standard deviation of the sample distribution within the same cluster in historical benchmark tests.

[0090] Retrieve all standard pressure response fingerprint samples that fall within the neighborhood of the hypersphere from the historical pattern library, count the number of valid samples and extract their corresponding feature coordinate data to form a local neighborhood sample subset.

[0091] Calculate the centroid coordinates of a subset of local neighborhood samples and evaluate the Euclidean distance between the current blister fingerprint vector to be tested and the local centroid to quantify the degree of deviation of the sample to be tested in the local topology.

[0092] A Gaussian kernel function is used to perform a weighted integral operation on the sample density distribution within the local neighborhood to construct a local consistency evaluation model. The local neighborhood consistency score is calculated using the following formula:

[0093]

[0094] Where S is the local neighborhood consistency score scalar, N is the number of valid samples in the neighborhood, vtest is the current blister fingerprint vector to be tested, vi is the i-th neighborhood sample vector, σ is the bandwidth parameter, and σ takes the value of the standard deviation of the distance between local neighborhood sample pairs.

[0095] The calculated local neighborhood consistency score is normalized to the [0,1] interval to generate a quantitative index characterizing the local clustering density of the current pressure response fingerprint in the feature space.

[0096] Through the above topological analysis and density-weighted calculation, the nearest neighbor cluster information determined in the previous step is transformed into a local neighborhood consistency score scalar that reflects the reliability of the sample distribution, thereby achieving a refined supplementary evaluation of the confidence level of the sealing state of the blister pack under test.

[0097] For example, the neighborhood radius R of the hypersphere is set to 0.8 times the average standard deviation within the cluster, and the number of valid samples N in the neighborhood is retrieved to be 15. The average Euclidean distance between the test vector and the neighborhood samples is obtained as 0.12, and the bandwidth parameter σ is set to 0.15. Substituting these values ​​into the formula to calculate the mean of the exponential term, the local neighborhood consistency score S is obtained as 0.92. This high score indicates that the pressure response pattern of the tested blister is highly consistent with the historical leak-free pattern, significantly improving the robustness of the discrimination results and effectively reducing the risk of misjudgment caused by a single distance metric.

[0098] S4.3: Obtain the minimum statistical distance value and the local neighborhood consistency score scalar in the multidimensional spatial statistical distance sequence, perform weighted fusion to linearly combine the reciprocal of the minimum statistical distance value and the local neighborhood consistency score scalar to comprehensively reflect the degree of proximity and distribution reliability of the test sample to the known leak-free mode, and generate the initial sealing state confidence index.

[0099] Obtain the minimum statistical distance value and the local neighborhood consistency score scalar in the multidimensional spatial statistical distance sequence as input parameters for weighted fusion.

[0100] Perform a reciprocal transformation operation on the minimum statistical distance value to construct a distance similarity mapping function, which transforms the Euclidean or Mahalanobis distance in the physical space into a similarity metric in the feature space. The smaller the distance, the higher the similarity.

[0101] A linear normalization algorithm is used to unify the dimensions of the distance similarity index after the reciprocal transformation, eliminating the interference of different feature dimensions and scale differences on the fusion weights, and generating standardized distance confidence components.

[0102] According to a preset weighting mechanism, weighting coefficients are calculated based on the dispersion of the current blister material type and historical pattern clusters, balancing the contribution ratios of distance similarity and neighborhood topology in the final judgment. The initial sealing state confidence is obtained by weighted summation of the reciprocal of the minimum statistical distance value and the local neighborhood consistency score. The weighting coefficients are preset values ​​that can be preset in the storage module according to different blister material types.

[0103] S4.4: Based on the initial sealing confidence index, a dynamic threshold adjustment function is invoked. Correction coefficients for material thickness and ambient temperature are extracted from the historical database according to the current blister pack's process label. These coefficients are then used to compensate for deviations in the initial sealing confidence index, adapting to the nonlinear response characteristics of blister packs of different specifications, and generating the final sealing confidence index. The temperature correction coefficient and thickness correction coefficient are empirical coefficients obtained through regression analysis of historical experimental data for different materials under different temperature and thickness conditions, and are pre-stored in the storage module.

[0104] The process label information associated with the current blister pack unit under test is analyzed to extract key physical property parameters, including packaging material type, nominal film thickness value and ambient temperature during testing.

[0105] Based on key physical property parameters, the historical benchmark test database is searched to locate the typical response mode cluster with the highest matching degree to the current operating conditions and its corresponding statistical distribution characteristic parameters.

[0106] The deviation of the current ambient temperature from the standard reference temperature is calculated, and a temperature-sensitive correction factor is derived by combining the material thermal expansion coefficient model to quantify the nonlinear impact of environmental fluctuations on the pressure response baseline.

[0107] Obtain the relative deviation rate between the measured and nominal values ​​of the current blister film thickness, and use the pre-stored thickness-stiffness coupling curve to query the mechanical response sensitivity correction coefficient to compensate for the mechanical deformation differences caused by material thickness tolerance.

[0108] A multidimensional deviation compensation mathematical model is constructed, and the temperature-sensitive correction factor and the mechanical response sensitivity correction coefficient are weighted and fused to generate a comprehensive correction weight vector.

[0109] The confidence index for the final sealing condition is calculated using the following formula:

[0110]

[0111] Wherein, Cfinal is the confidence index of the final sealing state, Cinit is the confidence index of the initial sealing state, α is the temperature correction coefficient, ΔT is the ambient temperature deviation value, β is the thickness correction coefficient, and ΔH is the relative deviation rate of film thickness.

[0112] The comprehensive correction weight vector is applied to the initial sealing state confidence index, and a linear deviation compensation operation is performed to eliminate systematic discrimination errors caused by process fluctuations.

[0113] Boundary truncation is performed on the compensated confidence index to ensure that its value range converges within the effective probability interval of 0 to 1, preventing logic overflow caused by extreme correction.

[0114] By using dynamic threshold adjustment and multi-physical quantity deviation compensation, the initial confidence level of the previous step is transformed into a final sealing state confidence index that adapts to specific process conditions, thereby achieving high-precision adaptation to the nonlinear response characteristics of blister packs of different specifications and significantly reducing the misjudgment rate.

[0115] For example, for a PVC / PVDC composite membrane blister with a nominal thickness of 0.25 mm, the testing environment temperature is 28℃ (standard 25℃). The extracted temperature deviation ΔT = 3℃ and membrane thickness deviation ΔH = 0.01 mm. From the table, the temperature correction factor α = -0.02 / ℃ and the thickness correction factor β = 0.5 / mm are obtained. If the initial confidence level Cinit is 0.85, substituting into the formula, we calculate: Cfinal = 0.85 × (1 + (-0.02 × 3) + (0.5 × 0.01)) = 0.85 × (1 - 0.06 + 0.005) = 0.85 × 0.945 = 0.80325. This process effectively compensates for the material softening caused by high temperature and the stiffness change caused by the slight increase in thickness, making the confidence index more accurately reflect the sealing state and significantly improving the accuracy of critical leak detection.

[0116] Step S5: Determine whether the sealing state confidence index meets the first judgment condition of falling into the known leak-free mode cluster and having a confidence level higher than a preset threshold. If the first judgment condition is not met, it is determined to be a critical response state and a perturbation verification command is triggered, generating a perturbation verification control signal. Specifically, this includes:

[0117] S5.1: Obtain the sealing state confidence index corresponding to the current blister unit under test and the pre-stored confidence threshold range of the no-leakage mode cluster. Based on the boundary inclusion detection algorithm, perform logical matching processing on whether the sealing state confidence index falls within the confidence threshold range of the no-leakage mode cluster to generate a Boolean interval matching result that characterizes the preliminary judgment of the sealing state.

[0118] S5.2: Based on the negative decision branch in the Boolean interval matching result, extract the outlier score of the pressure response fingerprint of the current blister unit under test in the feature space, and use the dynamic deviation calculation model to perform difference operation on the outlier score and the critical response judgment baseline to generate a critical response deviation metric that quantifies the uncertainty of the current sample.

[0119] The Boolean interval matching negative judgment result output by S5.1 is received, and the pressure response fingerprint of the current blister unit under test in the feature space is extracted. The Mahalanobis distance set between this fingerprint vector and the center of all known leak-free pattern clusters in the historical pattern library is calculated, and the minimum Mahalanobis distance is selected as the basic outlier index.

[0120] A dynamic bias calculation model is constructed, and a local density weighting coefficient is introduced to correct the baseline outlier. The local probability density of the fingerprint vector region is estimated using a Gaussian kernel function, and a density normalization factor is generated.

[0121] S5.3: Determine whether the current blister pack unit under test is in the fuzzy decision region between the known leak-free mode and the known leak mode based on the critical response deviation metric. If it is determined to be in the fuzzy decision region, mark it as a critical response state and generate a critical response state identifier to determine the set of target blister pack units that need to start the secondary verification mechanism.

[0122] Receive the critical response deviation metric generated in the previous step, which quantifies the degree of deviation of the current blister fingerprint vector under test relative to the boundary of the leak-free pattern cluster.

[0123] Set upper and lower boundary thresholds for the fuzzy decision region. The upper limit corresponds to the high-confidence, no-leakage state, and the lower limit corresponds to the explicit-leakage state. The area between the two constitutes the uncertainty interval.

[0124] The critical response deviation metric is compared with the upper and lower boundaries of the preset fuzzy decision region to determine whether it falls within the range.

[0125] If the metric value is between the upper and lower boundaries, the current blister is determined to be in a fuzzy decision region between the known leak-free mode and the known leak mode, and has mixed characteristics of micro-leakage or nonlinear material response.

[0126] For bubble cells identified as being in the fuzzy decision region, a unique critical response state identifier is generated, which includes a spatial coordinate index and an uncertainty level code.

[0127] Summarize the indices of all blister cells with critical response state identifiers to construct the target blister cell set that requires the activation of the secondary verification mechanism.

[0128] By using boundary comparison and state marking processing, the deviation metric value from the previous step is transformed into a structured target verification set, which enables precise locking of minor leakage risks under critical conditions and triggering of secondary verification, significantly reducing the false positive rate.

[0129] For example, the upper boundary threshold of the fuzzy decision region is set to 0.85, and the lower boundary threshold is set to 0.65. The critical response deviation metric of a certain blister cell is calculated to be 0.72. Since 0.65 < 0.72 < 0.85, the blister is determined to be within the fuzzy decision region. The system generates a critical response state identifier and adds it to the target blister cell set. This set is then passed to the perturbation verification module, triggering a micro-pressure pulse test for blister number 42, effectively avoiding missed detections or false alarms caused by fixed thresholds.

[0130] S5.4: For the set of target bubble cap units marked as critical response states, call the corresponding positive pressure pulse injection parameter configuration scheme in the perturbation excitation strategy library, and convert the positive pressure pulse injection parameter configuration scheme into a perturbation verification instruction data packet with specific timing characteristics based on the pulse width modulation coding rule, so as to define the subsequent action sequence of the vacuum proportional valve.

[0131] S5.5: Perform communication protocol encapsulation processing on the perturbation verification command data packet and send it to the vacuum control system of the sealed box to generate a perturbation verification control signal that drives the vacuum proportional valve to act, thereby triggering the physical process of injecting a small amount of controllable positive pressure pulse into the sealed box to complete the final leakage detection.

[0132] Step S6: Based on the perturbation verification control signal, inject a small amount of controllable positive pressure pulse into the sealed enclosure and simultaneously collect pressure recovery trajectory data. Analyze the spatiotemporal coupling characteristics of the pressure recovery trajectory data to identify local damping anomalies and generate the final leakage judgment result. Specifically, this includes:

[0133] S6.1: In response to the perturbation verification control signal, the vacuum proportional valve is pulse width modulated to inject a positive pressure pulse with an amplitude of 0.1 kPa to 0.3 kPa into the sealed housing, generating a positive pressure excitation airflow with a specific time domain width.

[0134] S6.2: The pressure recovery data of each bubble sensing unit under positive pressure excitation airflow is synchronously collected using a flexible pressure sensor array, and a sliding window differential operation is performed on the pressure recovery data to extract the pressure recovery slope data that characterizes the bubble deformation recovery rate.

[0135] S6.3: Based on the pressure recovery slope data, a four-dimensional spatiotemporal tensor containing spatial coordinate dimensions and time evolution dimensions is constructed, and a three-dimensional convolutional neural network is used to perform feature mapping transformation on the four-dimensional spatiotemporal tensor to generate a spatial gradient distribution map characterizing the pressure transmission hysteresis between adjacent blister packs.

[0136] The pressure recovery slope data sequence of each blister sensing unit output from S6.2 is received and combined with the physical space coordinate grid of the flexible sensor array to construct a four-dimensional spatiotemporal tensor data structure. The first and second dimensions of this tensor correspond to the row and column space coordinates of the sensor array, the third dimension is the time evolution axis, and the fourth dimension is the pressure recovery slope amplitude channel.

[0137] The constructed four-dimensional spatiotemporal tensor is subjected to feature mapping processing using a three-dimensional convolutional neural network. The network input layer receives tensor data and extracts gradient change features in the spatial neighborhood and dynamic response lag features on the time axis through three convolutional kernels.

[0138] S6.4: Calculate the damping coefficient deviation of the local area based on the spatial gradient distribution map, and compare the damping coefficient deviation with the preset critical damping threshold to identify the local damping abnormal area caused by gas micro-leakage and generate an abnormal location marker set.

[0139] The system receives the spatial gradient distribution map representing the pressure transmission hysteresis between adjacent bubble caps generated by S6.3, and extracts the local pressure recovery slope value and spatial coordinate information corresponding to each pixel in the map. Based on the fluid dynamics damping model, a local damping coefficient calculation equation is constructed, mapping the pressure recovery slope to an equivalent mechanical damping parameter. The local damping coefficient of the i-th sensing unit is calculated using the following formula:

[0140]

[0141] in, Let be the local damping coefficient of the i-th element. The amplitude of the positive pressure pulse. The stiffness coefficient of the blister material is derived by back-calculation from the standard damping coefficient of historical leak-free samples. This represents the pressure recovery slope.

[0142] A full-field damping coefficient matrix is ​​generated by traversing all nodes of the spatial gradient distribution map. The absolute value of the deviation between each local damping coefficient in this matrix and the corresponding standard damping mean in the historical leak-free model library is calculated to obtain a damping coefficient deviation sequence. A dynamic critical damping threshold is set, based on the standard damping mean and its three-times standard deviation, to cover the normal process fluctuation range. Each element in the damping coefficient deviation sequence is compared point-by-point with the critical damping threshold. Pixels with deviations exceeding the critical threshold are marked as potential leak points, and their spatial coordinates are recorded. Morphological dilation is performed on the marked points to connect adjacent abnormal pixels to form connected components, eliminating isolated noise points. The geometric center and coverage area of ​​the connected components are extracted to generate a set of abnormal location markers containing location information and severity. Through the above damping modeling and threshold comparison processing, the spatial gradient data from the previous step is transformed into a quantified leakage risk indicator, enabling precise location and identification of micro-leakage areas and significantly improving detection sensitivity. For example, for PVC blister packs, the positive pressure pulse amplitude is 0.2 kPa, and the material stiffness coefficient is 1.5 N / mm. The pressure recovery slope of a certain sensing unit is 0.05 kPa / s, and the calculated local damping coefficient is 2.67. The historical standard damping mean is 2.5, the standard deviation is 0.05, and the critical threshold is set to 2.65. Because 2.67 is greater than 2.65 and the deviation is significant, this point is marked as an anomaly. Through connected component analysis, a minor leak is confirmed, and the anomaly coordinates (12, 8) are output.

[0143] S6.5: Combine the set of abnormal location markers with the confidence index of the final sealing status to perform multi-source decision logic to output a final leakage judgment result that includes the state of good sealing, minor leakage, or serious leakage.

[0144] Obtain the damping coefficient deviation data of each blister unit in the abnormal location marker set, as well as the corresponding sealing status confidence index value.

[0145] A multi-source decision fusion model is constructed, which maps the damping coefficient deviation to a perturbation verification confidence component and the sealing state confidence index to a preliminary screening confidence component.

[0146] The weighted linear combination algorithm is used to calculate the comprehensive discrimination index, as shown in the following formula:

[0147]

[0148] Wherein, D is the comprehensive discriminant index, C is the confidence index of the sealed state, δ is the deviation of the normalized damping coefficient, and the weighting coefficients w1 and w2 are preset weight values, set according to the historical reliability ratio of the initial screening results and the perturbation verification results, and stored in the controller. Based on the comparison result between the comprehensive discriminant index D and the preset classification threshold range, the state classification logic decision is executed.

[0149] If D is greater than the high threshold, the blister is determined to be well sealed, and a normal status indicator is output.

[0150] If D is between the high threshold and the low threshold, it is determined that there is a minor leak, and the minor leak status indicator and leak level are output.

[0151] If D is less than the low threshold, it is determined that there is a serious leak or structural damage, and a serious leak status indicator is output.

[0152] By fusing multi-source data and using hierarchical threshold judgment, the dispersed pressure response characteristics and perturbation dynamic characteristics are transformed into a unified qualitative conclusion on the sealing state, thereby achieving high-precision and interference-resistant final leakage identification.

[0153] For example, weighting coefficients w1 and w2 are set to 0.4 and 0.6 respectively. The confidence index C of the sealing state of a certain blister unit is 0.75, and the deviation δ of the normalized damping coefficient is 0.3. Substituting these values ​​into the formula, the comprehensive discrimination index D is calculated as 0.4 * 0.75 + 0.6 * (1 - 0.3) = 0.3 + 0.4 * 2 = 0.72. The high threshold is set to 0.8, and the low threshold to 0.5. Since 0.5 < 0.72 < 0.8, the system determines that the blister has a minor leak. This method effectively distinguishes minor defects under critical conditions and significantly reduces the false positive rate.

[0154] Step S7: Update the sample distribution information in the historical pattern library based on the final leakage discrimination result. If an outlier fingerprint vector caused by a novel material or structure is detected, mark it as a seed point for a new category to optimize the partitioning boundary of subsequent typical response pattern clusters and generate dynamically updated adaptive discrimination model parameters. Specifically, this includes:

[0155] S7.1: Obtain the final leakage discrimination result and the corresponding pressure response fingerprint. Based on the outlier judgment label in the final leakage discrimination result, perform novelty evaluation processing on the pressure response fingerprint to generate a set of fingerprint vectors to be screened containing novelty scores.

[0156] S7.2: Perform density clustering boundary detection on the set of fingerprint vectors to be screened, and use the local neighborhood consistency score to calculate the Mahalanobis distance deviation between each vector and the center of the existing historical pattern cluster, so as to extract outlier fingerprint vectors that exceed the preset confidence interval as a potential new category candidate set.

[0157] S7.3: Based on the process batch and film thickness measurement labels associated with outlier fingerprint vectors in the potential new category candidate set, perform multidimensional feature space mapping transformation processing to convert the outlier fingerprint vectors into new category seed point data objects with clear physical attribute identifiers.

[0158] S7.4: Use the seed point data object of the new category to perform incremental update operation on the sample distribution information in the historical pattern library. Optimize the partition boundary of typical response pattern clusters by reconstructing the initial centroid set of the density peak clustering algorithm to generate an updated pattern library containing the topology of the new category.

[0159] S7.5: Recalculate the center position and covariance matrix parameters of each typical response pattern cluster based on the updated pattern library, and perform full synchronous update processing of the adaptive discriminant model parameters to output the latest adaptive discriminant model parameters with dynamic learning capabilities for the next round of detection tasks.

[0160] Example 2

[0161] Please refer to Figures 2 to 4 This embodiment proposes a sealing tester for executing the intelligent pressure gradient control method in Embodiment 1. The sealing tester includes:

[0162] The sealed enclosure 100 forms a closed space for accommodating the blister pack;

[0163] The vacuum assembly 200 is connected to the sealed chamber 100 and is used to evacuate the sealed chamber 100 and adjust the vacuum level inside the sealed chamber 100.

[0164] A flexible pressure sensor array 300 is disposed inside a sealed housing 100 and located above a blister pack, and has multiple sensing units arranged corresponding to each blister pack on the blister pack.

[0165] The lifting mechanism 400 is located inside the sealed housing 100 and is used to push the blister packs toward the flexible pressure sensor array 300 so that the top of each blister pack is pressed into contact with the flexible pressure sensor array 300.

[0166] Positioning stage 500 is set at the moving end of lifting mechanism 400 to fix blister packs and align each blister with the corresponding sensing unit in flexible pressure sensor array 300.

[0167] The controller (not shown in the figure) is electrically connected to the lifting mechanism 400, the flexible pressure sensor array 300 and the vacuum assembly 200 respectively, and is used to control their coordinated actions; wherein, the flexible pressure sensor array 300 is configured to detect the pressure changes at each sensing unit before and after the sealed box 100 is vacuumed, so as to reflect the sealing status of each blister point by point.

[0168] Specifically, the sealed chamber 100, as the core carrier, provides a sealed testing environment, ensuring the controllability of the vacuuming process and laying the environmental foundation for accurate pressure data acquisition. The sealed chamber 100 also features an openable and closable door. The vacuuming assembly 200, connected to the sealed chamber 100, enables precise adjustment of the vacuum level in the testing environment, allowing the system to simulate physical environments under different pressure gradients and induce differentiated pressure responses from the blister packs under different sealing conditions. The flexible pressure sensor array 300, arranged in a one-to-one correspondence with each blister pack on the blister pack plate, achieves real-time, spatial sensing of the pressure distribution on the blister pack surface.

[0169] Furthermore, the lifting mechanism 400 presses the blister pack plate into tight contact with the flexible pressure sensor array 300, ensuring effective mechanical coupling between the sensor and the blister pack, eliminating signal noise caused by poor contact, and guaranteeing the stability and consistency of pressure data acquisition. The positioning stage 500, by fixing and aligning the blister pack plate, ensures that each blister pack unit accurately corresponds to a specific sensing unit in the sensor array.

[0170] The flexible pressure sensor array 300 is configured to detect pressure changes before and after vacuuming, enabling the system to capture the dynamic evolution of the blister under negative pressure. By converting pressure changes into quantifiable data, it achieves an objective and quantitative assessment of the sealing status.

[0171] In some embodiments, the controller includes: a data acquisition module, including an analog-to-digital converter and a sampling clock, the data acquisition module being used to synchronously acquire voltage signals at each sensing unit of the flexible pressure sensor array 300 and transmit them to the processor via a communication interface; a processor, communicatively connected to the data acquisition module; and a storage module, communicatively connected to the processor, for storing a historical pattern library, process labels, and correction coefficients.

[0172] The data acquisition module is a key component used to convert the analog voltage signals output by the flexible pressure sensor array 300 into digital signals. The analog-to-digital converter (ADC) is responsible for performing the analog-to-digital conversion, and its accuracy and conversion rate directly affect the accuracy of subsequent data analysis. The sampling clock provides a precise time reference, ensuring that the ADC samples within a preset time interval, thereby guaranteeing the synchronization and consistency of data acquisition.

[0173] The core function of the data acquisition module is to achieve synchronous acquisition and efficient transmission of voltage signals from each sensing unit of the flexible pressure sensor array 300. The acquired digital signals are then transmitted to the processor via a communication interface.

[0174] The processor is the core computing unit that executes various algorithms in the intelligent pressure gradient control method. It receives digital pressure data from the data acquisition module and is responsible for performing complex operations such as normalization preprocessing, response feature extraction, pattern comparison, confidence calculation, and perturbation verification.

[0175] The storage module provides persistent storage for the processor, storing critical data required for the intelligent pressure gradient control method. The historical pattern library contains a large amount of known leak-free benchmark test data and their corresponding response pattern clusters, forming the basis for pattern comparison and sealing confidence calculation. Process tags identify the attributes of blister packs of different specifications, enabling the system to adapt to the characteristics of the blister pack under test. Correction coefficients compensate for deviations in the initial sealing confidence to accommodate the nonlinear responses of blister packs of different specifications.

[0176] The vacuum assembly 200 includes a vacuum pump 210 and a vacuum proportional valve 220. The vacuum pump 210 is connected to the sealed housing 100 and is used to evacuate the sealed housing 100; the vacuum proportional valve 220 is connected between the vacuum pump 210 and the sealed housing 100 and is used to adjust the vacuum level inside the sealed housing 100.

[0177] Vacuum pump 210 types include, but are not limited to, rotary vane vacuum pump 210, diaphragm pump, turbomolecular pump or Roots pump, etc., and can be selected according to factors such as the required vacuum range, pumping speed and cost.

[0178] The vacuum proportional valve 220 can be implemented using a proportional solenoid valve driven by a proportional electromagnet, a piezoelectric proportional valve driven by a piezoelectric ceramic, or a proportional valve combined with a mass flow controller (MFC).

[0179] A vacuum proportional valve 220 connects the vacuum pump 210 and the sealed housing 100, ensuring that the negative pressure generated by the vacuum pump 210 can act on the sealed housing 100 through the vacuum proportional valve 220. Specifically, the suction port of the vacuum pump 210 is connected to the inlet of the vacuum proportional valve 220 via a pipe, and the outlet of the vacuum proportional valve 220 is connected to the sealed housing 100 via another pipe. Sealed joints are typically used at the pipe connections to ensure the airtightness of the system. The controller sends a control signal to the vacuum proportional valve 220 to change the valve opening, thereby controlling the amount of air pumped into the sealed housing 100.

[0180] In some embodiments, the lifting mechanism 400 includes a servo motor 410, a transmission assembly 420, and a first support plate 430; the transmission assembly 420 includes a synchronous belt transmission structure 421, a ball screw 422, and a screw nut 423 threadedly engaged with the ball screw 422; the ball screw 422 is arranged vertically, and the screw nut 423 is fixedly installed on the first support plate 430; the servo motor 410 drives the ball screw 422 to rotate in the vertical direction via the synchronous belt transmission structure 421, and the ball screw 422 drives the first support plate 430 to move in the vertical direction via the screw nut 423, so that the positioning platform 500 disposed on the first support plate 430 rises and falls synchronously with the first support plate 430.

[0181] Specifically, the servo motor 410, as the power source of the lifting mechanism 400, provides precise speed and position control, responds quickly, and has high torque output capability. The transmission assembly 420 is a mechanical structure used to convert the rotational motion of the servo motor 410 into the linear lifting motion of the first support plate 430. Besides the combination of synchronous belt drive structure 421 and ball screw 422 used in this solution, the transmission assembly 420 can also be composed of various forms such as a rack and pinion mechanism, a chain drive mechanism, or a hydraulic / pneumatic cylinder. The first support plate 430 is used to fix the positioning stage 500 and support the blister pack.

[0182] In the transmission assembly 420, the synchronous belt drive structure smoothly transmits the rotational power of the servo motor 410 to the ball screw 422, which has the advantages of accurate transmission ratio, smooth operation, low noise, no need for lubrication, and effective absorption of instantaneous impact and vibration.

[0183] The ball screw 422 is arranged vertically to achieve the vertical lifting and lowering movement of the blister pack, ensuring a stable and perpendicular contact direction between the blister pack and the flexible pressure sensor array 300. The screw nut 423 is fixedly installed on the first support plate 430, ensuring that the driving force generated by the servo motor 410 through the ball screw 422 can directly and stably act on the first support plate 430, thereby driving it to perform precise vertical movement.

[0184] The lifting mechanism 400 operates as follows:

[0185] The controller sends a control command to the servo motor 410, which starts and outputs rotational power. The output shaft of the servo motor 410 transmits the rotational motion to the ball screw 422 via a synchronous belt drive structure, causing the ball screw 422 to rotate vertically. Since the screw nut 423 is fixedly mounted on the first support plate 430 and threadedly engaged with the ball screw 422, the rotational motion of the ball screw 422 is converted into linear movement of the first support plate 430 in the vertical direction via the screw nut 423. The first support plate 430 drives the positioning platform 500 fixed on it to move up and down synchronously, thereby causing the blister pack placed on the positioning platform 500 to move vertically closer to or further away from the flexible pressure sensor array 300.

[0186] When testing is required, the controller controls the servo motor 410 to rotate forward, the first support plate 430 moves upward, and the positioning stage 500 pushes the blister packs toward the flexible pressure sensor array 300 until the top of each blister pack is pressed into contact with the corresponding sensing unit. At this point, the servo motor 410 stops and remains locked in position to ensure stable pressing force. After testing is completed, the controller controls the servo motor 410 to rotate in the reverse direction, and the first support plate 430 moves downward, causing the blister packs to disengage from the flexible pressure sensor array 300, facilitating the removal of the blister packs and the placement of the next sample to be tested. Throughout the lifting process, the self-locking characteristic of the ball screw 422 ensures that the first support plate 430 remains stable at any stopping position, thereby guaranteeing the consistency and repeatability of contact between each blister pack and the sensing unit.

[0187] In some embodiments, the instrument also includes a frame 600; the frame 600 is equipped with a second support plate 700, which is arranged vertically opposite to the first support plate 430, and a flexible pressure sensor array 300 is mounted on the second support plate 700; the sealing tester also includes a guide rod 800 arranged vertically, which passes through the first support plate 430 and the second support plate 700 to guide the movement of the first support plate 430 relative to the second support plate 700; a buffer spring 900 is sleeved on the portion of the guide rod 800 located between the first support plate 430 and the second support plate 700 to buffer the impact when the first support plate 430 moves toward the second support plate 700.

[0188] Specifically, the frame 600 is the overall skeleton or support structure of the sealing tester, serving to support and fix the various functional modules. The second support plate 700 is a structural component for mounting the flexible pressure sensor array 300. It is arranged opposite to the first support plate 430, together forming the upper and lower limits and support of the lifting mechanism 400. The second support plate 700 is fixed to the frame 600 by bolts or clips. The guide rod 800 is a rod-shaped structure used to guide the first support plate 430 to move precisely in the vertical direction, ensuring the straightness and stability of the first support plate 430's movement trajectory. The guide rod 800 can be a precision-ground linear guide or optical shaft. The buffer spring 900 is provided on the guide rod 800 and is an elastic element used to absorb and disperse the mechanical impact energy generated when the first support plate 430 approaches the second support plate 700. It can be a helical compression spring.

[0189] The specific operating procedure of the sealing performance tester is as follows:

[0190] First, the blister pack to be tested is fixed on the positioning stage 500. The controller activates the lifting mechanism 400, driving the positioning stage 500 to move vertically toward the flexible pressure sensor array 300 until the top of each blister pack on the blister pack is pressed into contact with the corresponding sensing unit in the flexible pressure sensor array 300. Subsequently, the controller activates the vacuum assembly 200 to perform a vacuuming operation inside the sealed chamber 100.

[0191] Throughout the vacuuming process, the flexible pressure sensor array 300 continuously monitors the pressure change signals at each sensing unit. The data acquisition module synchronously samples the analog voltage signals and converts them into digital signals via an analog-to-digital converter, which are then transmitted to the processor via a communication interface. The processor executes the aforementioned intelligent pressure gradient control method.

[0192] Throughout the process, the controller coordinates the timing actions of the lifting mechanism 400, the vacuuming component 200, and the flexible pressure sensor array 300 according to the preset logic, and sequentially completes the pressing, vacuuming, signal acquisition, data discrimination, and micro-perturbation verification operations to achieve online, fully automatic, and non-destructive testing of the sealing status of medical blister packs.

[0193] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for intelligent control of pressure gradient, characterized in that, include: The pressure data collected by each bubble sensing unit in the flexible pressure sensor array during the vacuuming process is obtained, and the pressure data is normalized and preprocessed to generate relative pressure change data. Extract the response features of the relative pressure change data to generate the pressure response fingerprint of each blister unit; A historical pattern library is constructed based on historical leak-free benchmark test data. The historical pattern library is divided into several typical response pattern clusters by density peak clustering algorithm, the center position of each cluster is determined, and a set of historical pattern cluster centers with process labels is generated. The pressure response fingerprint is compared with the historical pattern library to calculate the sealing confidence of each blister unit. Determine whether the confidence level of the sealing state meets the preset judgment conditions; If the confidence level of the sealing state meets the preset judgment condition, then the corresponding blister unit is determined to be well sealed; If the confidence level of the sealing state does not meet the preset judgment condition, a perturbation verification is triggered and a perturbation verification control signal is generated. In response to the perturbation verification control signal, a positive pressure pulse is injected into the sealing chamber and pressure recovery data is collected. The final leakage judgment result of the corresponding blister unit is determined based on the pressure recovery data.

2. The intelligent pressure gradient control method according to claim 1, characterized in that, Also includes: The sample distribution information in the historical pattern library and the historical pattern cluster center set is updated based on the final leakage discrimination result. If an outlier fingerprint vector caused by a new material or structure is detected, it is marked as a seed point of a new category to optimize the division boundary of subsequent typical response pattern clusters and generate dynamically updated adaptive discrimination model parameters.

3. The intelligent pressure gradient control method according to claim 1, characterized in that, The steps for generating a stress response fingerprint include: Perform time-domain differential operation on the relative pressure change data sequence to obtain the initial contact slope parameter reflecting the mechanical properties at the instant of contact between the bubble and the sensor, and the maximum response delay time parameter characterizing the gas expansion response hysteresis effect. Based on the dynamic time window defined by the initial contact slope parameter and the maximum response delay time parameter, extreme value search and gradient accumulation calculation are performed on the rising phase data segment in the relative pressure change data sequence to generate the peak rise rate parameter describing the speed of blister inflating. Using the steady-state interval boundary determined by the peak rise rate parameter, the Shannon entropy estimation algorithm is applied to the plateau period data segment in the relative pressure change data sequence to quantify the degree of disorder in pressure fluctuations and generate plateau period fluctuation entropy parameters. By combining the initial contact slope parameter, the maximum response delay time parameter, the peak rise rate parameter, and the plateau period fluctuation entropy parameter, additional morphological feature parameters, including the position of the decay inflection point, are obtained to construct a multidimensional dimensionless feature set. A vector space mapping operation is performed on the multidimensional dimensionless feature set to encapsulate the dispersed feature parameters into a single bubble-cell pressure response fingerprint with a unique identifier, which serves as a standardized input object for subsequent clustering analysis and similarity comparison.

4. The intelligent pressure gradient control method according to claim 1, characterized in that, The steps for generating a confidence level for the sealed state include: Obtain the pressure response fingerprint generated by the current blister unit under test and the set of historical pattern cluster centers with process labels. Based on the Mahalanobis distance algorithm, perform covariance normalization on the Euclidean distance between the pressure response fingerprint and each cluster center in the set of historical pattern cluster centers to eliminate the influence of dimensional differences and correlations between different feature dimensions and generate a multidimensional spatial statistical distance sequence. Based on the multidimensional spatial statistical distance sequence, the target nearest neighbor cluster center is determined. Within a preset radius, the distribution density of sample points in the neighborhood of the target nearest neighbor cluster center is statistically analyzed to evaluate the local clustering degree of the current pressure response fingerprint in the feature space and generate a local neighborhood consistency score scalar. The minimum statistical distance value in the multidimensional spatial statistical distance sequence and the local neighborhood consistency score scalar are obtained. The reciprocal of the minimum statistical distance value and the local neighborhood consistency score scalar are weighted and summed according to a preset weight coefficient to generate an initial sealing state confidence index. Based on the initial sealing confidence index, a dynamic threshold adjustment function is invoked. The correction coefficients for the corresponding material thickness and ambient temperature are extracted from the historical database according to the process label of the current blister pack, and the initial sealing confidence index is subjected to deviation compensation processing to adapt to the nonlinear response characteristics of blister packs of different specifications, thereby generating the final sealing confidence index.

5. The intelligent pressure gradient control method according to claim 4, characterized in that, The steps to generate the final leak detection result include: In response to the perturbation verification control signal, the vacuum proportional valve is pulse width modulated to inject a positive pressure pulse with an amplitude of 0.1 kPa to 0.3 kPa into the sealed chamber. The flexible pressure sensor array is used to synchronously collect pressure recovery data of each bubble sensing unit under positive pressure excitation airflow, and a sliding window differential operation is performed on the pressure recovery data to extract pressure recovery slope data characterizing the bubble deformation recovery rate. Based on the pressure recovery slope data, a four-dimensional spatiotemporal tensor containing spatial coordinate dimensions and temporal evolution dimensions is constructed, and a three-dimensional convolutional neural network is used to perform feature mapping transformation on the four-dimensional spatiotemporal tensor to generate a spatial gradient distribution map characterizing the pressure transmission hysteresis between adjacent blister packs. The damping coefficient deviation of the local area is calculated based on the spatial gradient distribution map, and the damping coefficient deviation is compared and analyzed with the preset critical damping threshold to identify the local damping abnormal area caused by gas micro-leakage and generate an abnormal location marker set. The abnormal location marker set and the final sealing state confidence are combined to perform a multi-source decision logic judgment to output a final leakage discrimination result that includes whether the seal is intact or different leakage states.

6. A sealing tester for performing the intelligent pressure gradient control method according to any one of claims 1 to 5, characterized in that, include: The sealed enclosure forms a closed space for accommodating the blister pack; A vacuum pumping assembly, connected to the sealed enclosure, is used to evacuate the sealed enclosure and adjust the vacuum level inside the sealed enclosure. A flexible pressure sensor array is disposed inside the sealed box and above the blister pack, and has multiple sensing units arranged corresponding to each blister pack on the blister pack. A lifting mechanism, located inside the sealed box, is used to push the blister packs toward the flexible pressure sensor array, so that the top of each blister pack is pressed into contact with the flexible pressure sensor array. A positioning platform is set at the moving end of the lifting mechanism to fix the blister pack and align each blister with the corresponding sensing unit in the flexible pressure sensor array. The controller is electrically connected to the lifting mechanism, the flexible pressure sensor array, and the vacuum assembly, respectively, and is used to control their coordinated actions; wherein, the flexible pressure sensor array is configured to detect the pressure changes at each sensing unit before and after the sealed box is vacuumed, so as to reflect the sealing status of each blister point by point.

7. The sealing tester according to claim 6, characterized in that, The controller includes: The data acquisition module includes an analog-to-digital converter and a sampling clock. The data acquisition module is used to synchronously acquire the voltage signals at each sensing unit of the flexible pressure sensor array and transmit them to the processor via a communication interface. The processor is communicatively connected to the data acquisition module; The storage module is communicatively connected to the processor and is used to store historical pattern libraries, process tags, and correction coefficients.

8. The sealing tester according to claim 6, characterized in that, The vacuum pumping assembly includes: A vacuum pump, connected to the sealed housing, is used to evacuate the sealed housing. A vacuum proportional valve is connected between the vacuum pump and the sealed housing, and is used to adjust the vacuum level inside the sealed housing.

9. The sealing tester according to claim 6, characterized in that, The lifting mechanism includes a servo motor, a transmission assembly, and a first support plate; The transmission assembly includes a synchronous belt drive structure, a ball screw, and a screw nut that is threadedly engaged with the ball screw. The ball screw is arranged vertically, and the screw nut is fixedly installed on the first support plate; The servo motor drives the ball screw to rotate vertically via the synchronous belt transmission structure. The ball screw drives the first support plate to move vertically via the screw nut, so that the positioning platform set on the first support plate rises and falls synchronously with the first support plate.

10. The sealing tester according to claim 9, characterized in that, It also includes racks; The frame is equipped with a second support plate, which is arranged opposite to the first support plate in the vertical direction. The flexible pressure sensor array is mounted on the second support plate. The sealing tester also includes a guide rod arranged in a vertical direction, which passes through the first support plate and the second support plate to guide the movement of the first support plate relative to the second support plate; The portion of the guide rod located between the first support plate and the second support plate is fitted with a buffer spring to buffer the impact when the first support plate moves toward the second support plate.

Images