Infusion foreign matter online monitoring method and system based on unsupervised self-learning

Abstract

The application discloses an infusion foreign matter online monitoring method and system based on unsupervised self-learning, belongs to the technical field of infusion foreign matter online intelligent detection, and comprises the following steps: dividing a multi-source monitoring data set by screening core working condition dimensions and determining working condition stability weight coefficients, constructing an unsupervised correlation model and generating a correlation characteristic spectrum, establishing a double-branch collaborative output mechanism, outputting infusion foreign matter types and foreign matter type confidence by the first branch and determining recognition effectiveness, outputting infusion foreign matter prediction values under the current working condition by the second branch and determining prediction value effectiveness, calculating each working condition recognition baseline initial value to generate an initial recognition baseline, dynamically calibrating core parameters to generate a corrected infusion foreign matter recognition baseline, outputting online real-time recognition results by self-adaptive compensation signal deviation, calculating three-dimensional indexes, combining preset threshold values to construct three-dimensional constraint rules, determining infusion foreign matter monitoring risk grades and triggering processing mechanisms in stages, and optimizing the unsupervised correlation model, so that the infusion foreign matter recognition precision is improved.

Description

Technical Field

[0001] This invention relates to a method and system for online monitoring of foreign bodies in intravenous infusions based on unsupervised self-learning, belonging to the field of online intelligent detection technology for foreign bodies in intravenous infusions. Background Technology

[0002] Foreign body contamination is a significant factor affecting drug safety during the production and use of infusion preparations, especially in infusion products packaged in polypropylene plastic bottles. Although polypropylene material has good chemical stability and processing performance, it is prone to the shedding of tiny particles due to friction or environmental factors during injection molding, filling, and transportation, forming foreign bodies that can lead to clinical risks such as local tissue ischemia and necrosis, and thrombosis. Traditional methods for identifying foreign bodies in infusions mainly rely on manual interpretation or simple thresholds, which are difficult to adapt to the complex backgrounds and weak foreign body signals in online monitoring scenarios, and lack intelligent recognition capabilities for the morphology and material of foreign bodies. In addition, most existing foreign body monitoring technologies are based on supervised learning, requiring a large number of manually labeled foreign body samples, which is costly and has weak generalization ability, making it difficult to adapt to the dynamic changes in different production conditions and different infusion dosage forms during online monitoring.

[0003] A Chinese patent application with publication number CN118588234A discloses an intelligent infusion monitoring and early warning method and system, including: identifying infusion particles, establishing an intravenous infusion model to simulate extravasation, establishing a risk assessment model, real-time monitoring of infusion parameters, predicting risk levels, and sending an early warning command to the monitoring platform when an early warning is triggered. This addresses the technical problems of infusion monitoring relying on fixed threshold judgments, failing to adapt to different infusion conditions, and lacking adaptability to dynamic changes in infusion monitoring and early warning. The system achieves effective identification of potential risk factors during infusion through infusion particle identification and intravenous infusion simulation, combined with analysis of the physicochemical properties of drugs. It conducts multi-dimensional risk assessment of particle contamination and drug extravasation, dynamically adjusts risk thresholds and monitoring parameters, precisely controls infusion quality, adapts to different infusion conditions, and improves the effectiveness of infusion monitoring and early warning by combining real-time monitoring of infusion parameters.

[0004] Although there is an existing intelligent infusion monitoring and early warning method and system that identifies particles in infusion samples and combines them with intravenous infusion models for risk assessment, enabling dynamic monitoring and early warning of risk factors such as particulate contamination and drug extravasation during infusion, it does not consider the correlation between production parameters of polypropylene infusion bottles and the concentration and particle size distribution of foreign matter in infusions under various operating conditions. This results in the model being unable to adapt to the dynamic changes in the characteristics of foreign matter in infusions under online monitoring scenarios, leading to an increase in the false positive and false negative rates of foreign matter in infusions, and consequently, low accuracy in online identification of foreign matter in infusions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide an online monitoring method and system for infusion foreign bodies based on unsupervised self-learning. By selecting core operating conditions to determine stability weight coefficients to divide multi-source monitoring datasets, an unsupervised correlation model is constructed, and a dual-branch collaborative output mechanism and a dynamic baseline calibration mechanism are established. Combined with risk level determination and graded processing for infusion foreign bodies, the invention achieves high-precision, adaptive, and manual-label-free online identification and risk closed-loop management of infusion foreign bodies.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Online monitoring methods for foreign bodies in intravenous infusions based on unsupervised self-learning include:

[0008] Acquire multi-source monitoring datasets, filter core operating condition dimensions and determine operating condition stability weight coefficients, and divide the multi-source monitoring datasets.

[0009] An unsupervised association model is constructed to generate an association feature map. A two-branch collaborative output mechanism is established with real-time production parameters as input. The first branch outputs the type of infusion foreign body and the confidence level of the foreign body type to determine the identification effectiveness. The second branch outputs the predicted value of infusion foreign body under the current working conditions to determine the effectiveness of the predicted value.

[0010] Calculate the initial value of the identification baseline under each working condition, generate the initial identification baseline, dynamically calibrate the core parameters of online monitoring of infusion foreign bodies based on the effective predicted value of infusion foreign bodies, generate the corrected infusion foreign body identification baseline, adaptively compensate for signal deviation, and output the online real-time identification results;

[0011] The three-dimensional indicators are calculated and combined with preset thresholds to construct three-dimensional constraint rules, which are used to determine the risk level of foreign body monitoring in infusion and trigger a graded processing mechanism, and the unsupervised association model is iteratively optimized.

[0012] Specifically, the steps for dividing a multi-source monitoring dataset include:

[0013] Based on the multi-source monitoring dataset, injection molding temperature, blow molding pressure and sterilization time were selected as the core operating condition dimensions.

[0014] The analytic hierarchy process was used to determine the stability weight coefficients of the operating conditions, and the K-means clustering algorithm was used to cluster the core operating condition dimension data.

[0015] The absolute difference between the real-time values ​​of each core operating condition parameter and the industrial benchmark value is calculated. Based on the ratio of the absolute difference to the industrial benchmark value, and combined with the operating condition stability weighting coefficient, the comprehensive parameter fluctuation amplitude is calculated. ;

[0016] Set amplitude threshold , Divide the multi-source monitoring dataset;

[0017] like Then it is divided into a subset of stable operating conditions; if Then it is divided into a subset of fluctuating operating conditions; if Then it is divided into a subset of abnormal operating conditions;

[0018] Remove the abnormal operating condition subset and retain the stable operating condition subset and the fluctuating operating condition subset.

[0019] Specifically, the steps for generating the associated feature map include:

[0020] For the stable operating condition subset and the fluctuating operating condition subset, deep fusion features are extracted to generate a stable operating condition feature set and a fluctuating operating condition feature set;

[0021] Construct an unsupervised association model, including an input layer, a feature fusion enhancement layer, a dual-task processing layer, and an output layer;

[0022] The input layer receives the feature vectors of the stable operating condition feature set, the fluctuating operating condition feature set, and real-time production parameters;

[0023] The feature fusion enhancement layer performs dimensional restructuring and correlation enhancement on the deep fusion features, and focuses on the strongly correlated features of production parameters and infusion foreign body characteristics through an attention mechanism to establish a correlation weight matrix;

[0024] The dual-task processing layer utilizes a lightweight graph convolutional network in the type recognition branch to learn the mapping relationship between the fluctuation of working condition parameters and the type of foreign object and extract foreign object features based on the association weight matrix.

[0025] The quantitative prediction branch uses a fully connected regression network to simultaneously learn the quantitative correlation between production parameters and foreign matter concentration and particle size distribution, and outputs a predicted feature vector.

[0026] Specifically, the steps for generating the associated feature map also include:

[0027] The output layer establishes a dual-branch collaborative output mechanism. The first branch outputs the type of foreign body in the infusion and the confidence level of the foreign body type, while the second branch outputs the predicted value of the foreign body in the infusion, including the predicted value of particle size distribution and the predicted value of foreign body concentration.

[0028] The stable operating condition feature set and the fluctuating operating condition feature set are divided into a training set and a validation set according to a preset ratio;

[0029] The unsupervised correlation model is trained based on the training set, using a joint loss function and the Adamw optimizer.

[0030] After training, the performance of the unsupervised association model is verified using a validation set. The recognition accuracy and prediction mean square error of foreign object types are calculated, and accuracy thresholds and error thresholds are set.

[0031] If the recognition accuracy is greater than or equal to the accuracy threshold and the prediction mean square error is less than or equal to the error threshold, the unsupervised association model is deemed to have met the standard. After meeting the standard, an association feature map is generated based on the association weight matrix.

[0032] Otherwise, adjust the optimizer learning rate and retrain.

[0033] Specifically, the steps for determining the validity of predicted values ​​include:

[0034] The real-time production parameters are input into the unsupervised association model after training has reached the target. The first branch matches the type of foreign object and the confidence level of the foreign object type based on the association feature map.

[0035] Set a confidence threshold to determine the effectiveness of foreign object type identification;

[0036] If the confidence level of the foreign object type is greater than or equal to the confidence threshold, the foreign object type identification result is deemed valid; otherwise, a model warning is triggered and the current production parameter fluctuation information is recorded.

[0037] The second branch outputs the predicted particle size distribution and the predicted foreign matter concentration under the current operating conditions;

[0038] Set prediction thresholds, including a first prediction threshold and a second prediction threshold, to determine the validity of the predicted values;

[0039] If the predicted particle size distribution is less than or equal to the first prediction threshold and the predicted foreign body concentration is less than or equal to the second prediction threshold, then the predicted foreign body concentration in the infusion is deemed valid.

[0040] Otherwise, the predicted value of foreign matter in the infusion is deemed invalid, the corresponding infusion product is marked as unqualified, and the production line is activated to perform diversion processing.

[0041] Specifically, the steps for generating the initial identification baseline include:

[0042] Using the infusion foreign body characteristic data within a unified historical optimal monitoring period as a benchmark, cluster centers for each working condition are extracted from the associated feature map.

[0043] The visualized association information of cluster centers is quantified into cluster feature vectors to form a cluster center feature set for each working condition;

[0044] Calculate the correlation strength between the characteristics of each cluster center and the properties of infusion foreign bodies;

[0045] If the correlation strength is greater than or equal to the preset screening threshold, the clustering feature is deemed valid and saved to the cluster center feature set of the corresponding working condition; otherwise, invalid clustering features are removed.

[0046] For each working condition, calculate the mean of the best foreign object characteristics in history, extract the effective cluster center foreign object characteristic values, and calculate the first and second baseline initial values ​​of foreign object characteristics by weighting.

[0047] Set integration weights, and integrate the initial values ​​of the first and second baselines to form the initial particle size identification baseline and the initial concentration identification baseline. Then, merge them according to the foreign matter characteristic dimension to generate the initial identification baseline.

[0048] Specifically, the steps for generating a corrected baseline for foreign body identification in infusions include:

[0049] Based on effective infusion foreign body prediction values, core parameters for online infusion foreign body monitoring are calibrated in multiple dimensions.

[0050] In terms of exposure intensity, the near-infrared spectrum of the polypropylene infusion bottle is obtained, and the transmittance is estimated in real time based on the correspondence between the near-infrared spectrum and transmittance in order to calibrate the exposure intensity.

[0051] In terms of signal acquisition frequency, adjustments are made in stages based on the amplitude of fluctuations in operating conditions.

[0052] If the fluctuation range of the operating condition is less than or equal to the preset fluctuation threshold, the basic signal acquisition frequency is maintained; otherwise, the signal acquisition frequency is increased in stages.

[0053] The pre-signal for foreign object detection is acquired in real time using the calibrated core parameters, and the deviation rate between the foreign object particle size distribution in the pre-signal and the predicted particle size distribution is calculated.

[0054] If the deviation rate is less than or equal to the preset deviation limit, the fluctuation range is expanded based on the predicted particle size distribution to form a particle size correction baseline;

[0055] Otherwise, based on historical deviation correction data under different operating conditions, the fluctuation range of the initial concentration identification baseline is expanded and the particle size correction baseline is recalculated;

[0056] Finally, a corrected baseline for foreign body identification in infusion is generated.

[0057] Specifically, the steps for adaptive compensation of signal deviation include:

[0058] Signal deviation is compensated step by step using an adaptive compensation algorithm;

[0059] For differences in material transmittance, the standard transmittance is obtained and the real-time estimated transmittance is combined to calculate the compensation coefficient. The gain of the original signal amplitude is adjusted to obtain the material deviation compensation signal.

[0060] For the deviation of operating conditions, the interference component that matches the characteristics of the operating conditions fluctuation is extracted from the original signal, and the deviation between the signal phase and the standard phase of the corrected infusion foreign body identification baseline is reduced to obtain the operating condition fluctuation deviation compensation signal.

[0061] The material deviation compensation signal and the working condition fluctuation deviation compensation signal are fused to form a full compensation signal, which is then compared feature by feature with the corrected infusion foreign body identification baseline to output the online real-time identification result of infusion foreign body;

[0062] If the particle size and concentration of foreign matter in the full compensation signal are within the threshold range corresponding to the corrected infusion foreign matter identification baseline and the signal-to-noise ratio is greater than or equal to the industry qualified threshold, then the online identification result is determined to be normal.

[0063] Otherwise, the online recognition result will be deemed abnormal.

[0064] Specifically, the steps for determining the risk level of foreign body monitoring during intravenous infusion include:

[0065] By comparing the online real-time identification results with the prediction results of the unsupervised association model, the foreign matter concentration deviation value, particle size matching degree, and confidence consistency index are calculated.

[0066] Set the concentration deviation threshold, particle size matching threshold, and confidence consistency threshold;

[0067] Construct three-dimensional constraint rules and combine them with real-time online identification results to determine the risk level of foreign body monitoring in infusion.

[0068] When the first constraint rule is met and the online recognition result is normal, it is judged as a level 1 risk;

[0069] When the second constraint rule is met and both particle size and concentration exceed the standard, it is judged as a level 2 risk.

[0070] When the third constraint rule is met and there are double exceedances, misjudgments, and missed detections, it is judged as a level three risk;

[0071] After classifying and processing risks at each level, operating condition features and indicator deviation features are extracted for the secondary and tertiary risk samples and input into the unsupervised association model for incremental learning.

[0072] The core parameters of dynamic calibration are optimized based on three levels of risk samples, and the unsupervised association model is iteratively optimized.

[0073] The online monitoring system for foreign bodies in infusion based on unsupervised self-learning includes: a multi-source fusion module, a correlation prediction module, a baseline dynamic calibration module, and a dual-source comparison module;

[0074] The multi-source fusion module is used to collect multi-dimensional production parameter data and infusion foreign matter characteristic data of polypropylene infusion bottles under different production conditions in real time, and generate multi-source monitoring datasets after preprocessing.

[0075] The correlation prediction module is used to filter core working condition dimensions and determine the working condition stability weight coefficient, divide the multi-source monitoring dataset; construct an unsupervised correlation model, generate a correlation feature map, input real-time production parameters, and establish a dual-branch collaborative output mechanism. The first branch outputs the infusion foreign body type and the foreign body type confidence level to determine the identification effectiveness, and the second branch outputs the infusion foreign body prediction value under the current working condition to determine the prediction value effectiveness.

[0076] The baseline dynamic calibration module is used to calculate the initial value of the identification baseline under each working condition, generate the initial identification baseline, dynamically calibrate the core parameters of online monitoring of infusion foreign bodies based on the effective predicted value of infusion foreign bodies, generate the corrected infusion foreign body identification baseline based on the predicted value of particle size distribution, adaptively compensate for signal deviation, and output the online real-time identification result of infusion foreign bodies.

[0077] The dual-source comparison module is used to calculate three-dimensional indicators, construct three-dimensional constraint rules in combination with preset thresholds, determine the risk level of foreign body monitoring in infusion and trigger a graded processing mechanism, and iteratively optimize the unsupervised association model.

[0078] The beneficial effects of this invention are:

[0079] 1. By selecting core operating condition dimensions to determine stability weight coefficients, and combining K-means clustering to divide stable and fluctuating operating condition subsets, an unsupervised association model is constructed. Through an attention mechanism, the strong correlation features between production parameters and foreign object characteristics are focused, and a dual-branch collaborative output mechanism is established to simultaneously output the type of infusion foreign object, the confidence level of the foreign object type, and the predicted value of infusion foreign object and to determine its effectiveness. Based on near-infrared spectroscopy, the transmittance is estimated in real time to calibrate the exposure intensity. The signal acquisition frequency is adjusted according to the level of operating condition fluctuations, and the initial identification baseline is dynamically corrected to generate a corrected infusion foreign object identification baseline. The model adaptively compensates for the signal deviation caused by differences in material transmittance and fluctuations in operating conditions, effectively solving the problem of insufficient adaptability of traditional models due to dynamic changes in operating conditions, reducing the false positive rate and false negative rate of foreign objects and improving the accuracy of identification results.

[0080] 2. An unsupervised association model is constructed through unsupervised self-learning. Training and validation are completed by combining the joint loss function and Adamw optimizer, and the identification and prediction effectiveness are determined. The production line is then linked to perform diversion processing for non-conforming products. By calculating three-dimensional indicators of concentration deviation, particle size matching degree, and confidence consistency, three-dimensional constraint rules are constructed for risk classification. The working condition characteristics and indicator deviation characteristics of level 2 and level 3 high-risk samples are fed back to the unsupervised association model for incremental learning. Based on level 3 risk samples, the core parameters are dynamically calibrated, effectively reducing labor costs and meeting the real-time requirements of online monitoring in infusion production while improving quality control efficiency, thus providing strong protection for the quality and safety of infusion products. Attached Figure Description

[0081] Figure 1 This is a schematic diagram of an online monitoring method for foreign bodies in intravenous infusions based on unsupervised self-learning.

[0082] Figure 2 This is a flowchart for determining the validity of predicted values ​​in this invention;

[0083] Figure 3 This is a flowchart illustrating the process of generating a corrected infusion foreign body identification baseline in this invention;

[0084] Figure 4 This is a flowchart illustrating the risk level determination for foreign body monitoring during intravenous infusion in this invention;

[0085] Figure 5 This is a structural diagram of the online monitoring system for foreign bodies in infusion based on unsupervised self-learning in this invention. Detailed Implementation

[0086] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0087] Example 1

[0088] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces an online monitoring method for foreign bodies in intravenous infusions based on unsupervised self-learning, including the following steps:

[0089] Multi-dimensional production parameter data of polypropylene infusion bottles under different production conditions are collected in real time by a multi-sensor array, such as injection pressure and blow molding time. Foreign matter characteristic data of infusion samples under corresponding conditions are collected by a high-precision particle detection device, such as foreign matter concentration and particle size distribution. Gaussian filtering is used to remove sensor noise, and time stamp alignment is used to match the production parameters and foreign matter characteristic data under different conditions. Z-score standardization is used to eliminate dimensional differences, box plot method is used to remove outlier data, and principal component analysis is used to reduce dimensionality and remove data redundancy, generating a standardized multi-source monitoring dataset. Among them, outlier data includes, but is not limited to, data of false triggering by detection equipment.

[0090] Based on a multi-source monitoring dataset, injection temperature, blow molding pressure, and sterilization time were selected as core operating condition dimensions. The analytic hierarchy process (AHP) was used to determine the stability weight coefficients for each operating condition, and K-means clustering was combined to partition the multi-source monitoring dataset. An unsupervised autoencoder was used to extract deep fusion features between production parameters and infusion foreign body characteristics, constructing an unsupervised correlation model. This model autonomously learns and fits the dynamic correlation patterns between production parameters and infusion foreign body characteristics under various operating conditions. The Force-Atlas2 layout algorithm was used to generate a correlation feature map, visually presenting the correlation between fluctuations in production parameters and changes in infusion foreign body characteristics. Real-time production parameters were input into the unsupervised correlation model, establishing a dual-branch collaborative output mechanism. The first branch outputs the infusion foreign body type and its confidence level, determining the effectiveness of foreign body type identification. The second branch outputs the predicted values ​​of infusion foreign bodies under the current operating condition, including predicted particle size distribution and foreign body concentration, determining the validity of the predicted values ​​and removing predicted data corresponding to invalid infusion products.

[0091] Based on the infusion foreign body characteristic data within the historical best monitoring period, and combined with the clustering centers of the correlation feature maps under various operating conditions, the initial value of the identification baseline under each operating condition is calculated by weighting the features of the clustering centers and the mean of the historical best data, thus generating an initial identification baseline. Based on the effective predicted value of infusion foreign bodies, a dynamic calibration mechanism for the infusion foreign body identification parameters is constructed to dynamically calibrate the core parameters of online monitoring of infusion foreign bodies. A corrected infusion foreign body identification baseline is generated based on the predicted value of particle size distribution. An adaptive compensation algorithm adaptively compensates for signal deviations, outputting the online real-time identification result of infusion foreign bodies. The core parameters include, but are not limited to, the exposure intensity of the detection equipment and the signal acquisition frequency. The signal deviation is caused by the difference in the light transmittance of the polypropylene infusion bottle material and the fluctuation of operating conditions.

[0092] The online real-time identification results of foreign bodies in infusions are compared with the prediction results of unsupervised association models. Three-dimensional indicators are calculated, including foreign body concentration deviation, particle size matching degree, and identification confidence consistency index. Three-dimensional constraint rules are constructed in combination with preset thresholds to determine the risk level of foreign body monitoring in infusions and trigger a graded processing mechanism. The unsupervised association model and calibration mechanism are iteratively optimized. Among them, the preset thresholds include concentration deviation threshold, matching degree threshold, and confidence consistency threshold.

[0093] Specifically, the steps for dividing a multi-source monitoring dataset include:

[0094] Based on a multi-source monitoring dataset, injection temperature, blow molding pressure, and sterilization time were selected as core operating condition dimensions in the production process of polypropylene infusion bottles. According to historical operational stability data of the polypropylene infusion bottle production equipment, the analytic hierarchy process (AHP) was used to determine the stability weight coefficients for each operating condition. K-means clustering was then used to cluster the core operating condition dimension data. Specifically, the AHP constructed a hierarchical structure consisting of a production operating condition stability target layer, a core operating condition dimension criterion layer, and a stability influencing factor scheme layer. Industry experts scored the stability impact of each single dimension in the core operating condition dimension criterion layer using pairwise comparisons, constructed a judgment matrix, and calculated the maximum eigenvalue of the judgment matrix to obtain the stability weight coefficients for each single dimension. The K-means clustering algorithm standardized the raw data of injection temperature, blow molding pressure, and sterilization time extracted from the multi-source monitoring dataset. Based on the stability weight coefficients, the standardized data were weighted and fused to generate a comprehensive operating condition feature vector. Using this comprehensive feature vector as input, cluster centers were initialized, and the Euclidean distance from each sample to the cluster center was iteratively calculated. Samples were assigned to the nearest cluster, and the cluster centers were updated until the cluster centers remained unchanged.

[0095] Temperature sensors, pressure sensors, and sterilization timers deployed on polypropylene infusion bottle production equipment are used to collect injection temperature, blow molding pressure, and sterilization time in real time as core operating condition parameters. An industry benchmark value is determined by weighting the rated parameters specified in the polypropylene infusion bottle production process standard, the equipment's factory calibration values, and the historical average of the best operating condition parameters. The absolute difference between each core operating condition parameter value and the industry benchmark value is calculated, and the ratio of this absolute difference to the industry benchmark value is used to obtain a one-dimensional fluctuation coefficient. This one-dimensional fluctuation coefficient is multiplied by the corresponding operating condition stability weighting coefficient to obtain a one-dimensional weighted fluctuation value. The summation of these one-dimensional weighted fluctuation values ​​yields the comprehensive parameter fluctuation amplitude. ;

[0096] The amplitude threshold is set based on historical best production cycle parameter fluctuation statistics and industry expert experience. , Divide the multi-source monitoring dataset;

[0097] like Then it is divided into a subset of stable operating conditions; if Then it is divided into a subset of fluctuating operating conditions; if If the condition is abnormal, it is divided into an abnormal operating condition subset; the abnormal operating condition subset is removed and the stable operating condition subset and the fluctuating operating condition subset are retained.

[0098] Specifically, the steps for determining the validity of predicted values ​​include:

[0099] For the stable operating condition subset and the fluctuating operating condition subset, an unsupervised autoencoder is used to extract deep fusion features of production parameters and infusion foreign matter characteristics, such as parameter fluctuation trend features and foreign matter concentration correlation features, to generate stable operating condition feature sets and fluctuating operating condition feature sets.

[0100] Based on the dual-branch collaborative output requirement and the unsupervised association learning logic, an unsupervised association model is constructed, including an input layer, a feature fusion enhancement layer, a dual-task processing layer, and an output layer.

[0101] The input layer is used to receive feature vectors of stable operating condition feature sets, fluctuating operating condition feature sets, and real-time production parameters;

[0102] The feature fusion enhancement layer is used to reorganize and strengthen the association of deep fusion features. It focuses on the strong correlation features between production parameters and infusion foreign body characteristics through the attention mechanism and establishes a correlation weight matrix. The matrix contains the correlation strength values ​​between each core working condition parameter and infusion foreign body characteristics.

[0103] The dual-task processing layer consists of a type recognition branch and a quantization prediction branch. The type recognition branch uses a lightweight graph convolutional network to learn the mapping relationship between the fluctuation of operating parameters and the type of foreign matter based on the correlation weight matrix, and extracts foreign matter features, such as plastic debris and fibers. The quantization prediction branch uses a fully connected regression network to simultaneously learn the quantization correlation law between production parameters and foreign matter concentration and particle size distribution, and outputs a continuous prediction feature vector.

[0104] The output layer is used to establish a dual-branch collaborative output mechanism. The first branch outputs the type of foreign body in the infusion and the confidence level of the foreign body type, and the second branch outputs the predicted value of the foreign body in the infusion, including the predicted value of particle size distribution and the predicted value of foreign body concentration.

[0105] The stable operating condition feature set and the fluctuating operating condition feature set are divided into training set and validation set according to a preset ratio. Based on the training set, an unsupervised correlation model is trained by combining the joint loss function and the Adamw optimizer. In this embodiment, the preset ratio is set to 8:2. The joint loss function is determined by combining the comparison loss function, the cross-entropy loss function and the mean squared error loss function.

[0106] After training, the performance of the unsupervised association model is validated using a validation set. The model predictions on the validation set are compared with the actual foreign body characteristic data. The prediction results of all samples in the validation set are statistically analyzed. The category with the highest confidence in the foreign body type output by the unsupervised association model is matched with the actual foreign body type, and the proportion of correctly matched samples to the total number of samples in the validation set is calculated to obtain the foreign body type identification accuracy. For the infusion foreign body prediction value output by the second branch, the total number of samples in the validation set is statistically analyzed. and the number of quantiles contained in the particle size distribution of each sample. For each sample, calculate the square of the concentration difference between the predicted and actual foreign body concentrations. The squared difference between the predicted and actual particle size quantile values ​​is calculated by summing the squared concentration difference for each sample with the squared particle size difference, and then adding the squared concentration difference to the squared particle size difference to obtain the total squared error for a single sample. The sum of the squared errors of each sample is the total sum of squared errors. The total sum of squared errors is then divided by the total number of data points. The prediction mean square error is obtained; wherein, in this embodiment Indicates the concentration dimension;

[0107] Accuracy and error thresholds are set based on industry technical standards and historical best-test data. If the recognition accuracy is greater than or equal to the accuracy threshold and the prediction mean square error is less than or equal to the error threshold, the unsupervised association model is deemed to have met the standard. After meeting the standard, a visualized association feature map is generated based on the association weight matrix using the Force-Atlas2 layout algorithm; otherwise, the optimizer learning rate is adjusted and the model is retrained. In the association feature map, nodes represent working condition parameters and foreign object characteristics, and edges represent association strength.

[0108] The unsupervised correlation model, trained to meet the training criteria, is fed real-time production parameters. Combined with a two-branch collaborative output mechanism in the output layer, the first branch matches foreign object types and their confidence levels based on hotspot regions of the correlation feature map. A confidence threshold is set based on industry expert experience to determine the validity of foreign object type identification. If the foreign object type confidence level is greater than or equal to the confidence threshold, the identification result is considered valid; otherwise, it is deemed invalid, triggering a model warning and recording current production parameter fluctuations. The second branch outputs predicted particle size distribution and foreign object concentration under the current operating conditions. Prediction thresholds, including a first and a second prediction threshold, are set according to the infusion foreign object limit standard to determine the validity of the predicted values. If the predicted particle size distribution is less than or equal to the first prediction threshold and the predicted foreign object concentration is less than or equal to the second prediction threshold, the predicted infusion foreign object value is considered valid; otherwise, it is deemed invalid, the corresponding infusion product is marked as non-conforming, and the production line is instructed to perform diversion processing.

[0109] For example, taking the infusion production scenario as an example, the input layer of the unsupervised association model receives a stable operating condition feature set, a fluctuating operating condition feature set, and feature vectors of real-time production parameters; wherein the stable operating condition feature set includes the average infusion flow rate of 2.5 m / s, the temperature of 35℃, and the pipeline material of PVC, and the fluctuating operating condition feature set includes the standard deviation of pressure fluctuation of 0.3 kPa and the peak value of flow rate fluctuation of 0.8 m / s;

[0110] The feature fusion layer focuses on strongly correlated features through an attention mechanism to establish a correlation weight matrix; among them, the correlation strength between infusion flow rate and plastic debris is 0.88, the correlation strength between pressure fluctuation and fiber is 0.75, and the correlation strength between temperature and foreign matter concentration is set to 0.69.

[0111] The 800 sets of stable operating condition feature sets and fluctuating operating condition feature sets were divided into 640 training sets and 160 validation sets according to a preset ratio of 8:2. An unsupervised correlation model was trained using an Adamw optimizer with an initial learning rate of 0.001 combined with a joint loss function. The output results after training, for example, when the input feature vector has a flow rate of 3.0 ml / s and a pressure fluctuation of 0.4 kPa, show that the first branch outputs plastic debris with a confidence level of 0.94; the second branch outputs a particle size distribution of 70% (0.05-0.1 mm) and 25% (0.1-0.2 mm), with a predicted foreign matter concentration of 3. 0.2 particles / 100ml; the foreign object type identification accuracy was 93% and the prediction mean square error was 0.03, which was compared with the industry standard of 85% accuracy and 0.05 error, indicating that the unsupervised association model met the standard. Based on the association weight matrix, the Force-Atlas2 algorithm was used to generate a visualization map, with infusion flow rate and plastic debris as nodes, and the thickness of the edges corresponding to the association strength. For example, the edges of infusion flow rate and plastic debris were 0.88 times the baseline. Among them, the contrast loss weight was set to 0.3, the cross-entropy loss weight was set to 0.4, and the mean square error loss weight was set to 0.3.

[0112] Specifically, the steps for generating the initial identification baseline include:

[0113] The data on foreign body characteristics in infusion during the unified historical optimal monitoring period are obtained from the multi-source monitoring historical database and used as a benchmark. Cluster centers for each working condition are extracted from the correlation feature map. The feature quantization method is used to quantify the visualized correlation information of the cluster centers into cluster feature vectors, such as the corresponding values ​​of foreign body characteristics, to form a cluster center feature set for each working condition, including the cluster center feature set for stable working conditions and the cluster center feature set for fluctuating working conditions.

[0114] Corresponding sample data pairs of cluster center features and infusion foreign body characteristics were selected. The Pearson correlation coefficient was used to calculate the linear correlation coefficient between each cluster center feature and the infusion foreign body characteristics. The absolute value of the linear correlation coefficient was taken to obtain the correlation strength. Feature validity was screened. A screening threshold was set based on the production process characteristics of polypropylene infusion bottles and the experience of industry experts. If the correlation strength was greater than or equal to the screening threshold, the cluster feature was deemed valid and saved to the cluster center feature set of the corresponding working condition; otherwise, the cluster feature was deemed invalid and invalid cluster features were removed. Among them, the cluster center features include the mean infusion flow rate, the standard deviation of pressure fluctuation, and the variance of temperature fluctuation.

[0115] For each operating condition, the average value of the best foreign matter characteristics in history is calculated using the arithmetic mean method. Effective foreign matter characteristic values ​​of cluster centers are extracted from the cluster center feature set. The first and second baseline initial values ​​of foreign matter characteristics are calculated by weighting. Among them, foreign matter characteristics include particle size and concentration. Each operating condition includes stable operating condition and fluctuating operating condition.

[0116] For example, the effective foreign matter concentration samples under stable operating conditions extracted from the historical database are [2.1, 2.3, 2.0, 2.2], with a total concentration of 2.1 + 2.3 + 2.0 + 2.2 = 8.6, and the average of the highest foreign matter concentration under stable operating conditions is 8.6 / 4 = 2.15;

[0117] Based on the frequency of actual production under stable and fluctuating operating conditions, the integration weights are set, and the initial values ​​of the first and second baselines are weighted and integrated to form the initial identification baseline for particle size and the initial identification baseline for concentration. They are then merged according to the foreign matter characteristics dimension to generate the initial identification baseline for infusion foreign matter.

[0118] Specifically, the steps for adaptive compensation of signal deviation include:

[0119] Based on effective infusion foreign body prediction values, a dynamic calibration mechanism for infusion foreign body identification parameters is constructed. The core parameters of online infusion foreign body monitoring are calibrated in multiple dimensions. In the dimension of exposure intensity, each polypropylene infusion bottle is rapidly scanned by an online near-infrared spectral sensor to obtain the near-infrared spectrum of the polypropylene infusion bottle. The transmittance is estimated in real time using a correspondence table between near-infrared spectrum and transmittance based on experimental data of the spectral characteristics of polypropylene material. Then, based on the real-time transmittance estimation, the exposure intensity is calibrated using the mapping relationship between transmittance and exposure intensity fitted based on historical data. For example, when the transmittance decreases, the exposure intensity is increased to compensate for signal attenuation.

[0120] In terms of signal acquisition frequency, adjustments are made in stages based on the amplitude of fluctuations in operating conditions. A fluctuation threshold is set based on the baseline values ​​of production parameters under stable operating conditions and the allowable fluctuation range of the polypropylene infusion bottle production process. If the fluctuation amplitude is less than or equal to the fluctuation threshold, the basic signal acquisition frequency is maintained; otherwise, the signal acquisition frequency is increased in stages according to the fluctuation amplitude. For example, if the fluctuation amplitude is greater than 10%, the signal acquisition frequency is increased to 120Hz. The fluctuation amplitude of operating conditions is determined by the weighted relative deviation between the actual values ​​of production parameters acquired in real time and the industrial baseline values. In this embodiment, the basic signal acquisition frequency is set to 50Hz.

[0121] A set of pre-signals for foreign matter detection is acquired in real time using calibrated core parameters. Based on the absolute difference between the foreign matter particle size distribution and the predicted particle size distribution value in the pre-signal, the ratio of the absolute difference to the predicted particle size distribution value is calculated to obtain the deviation rate. Deviation limits are set according to the industry standard for foreign matter detection in polypropylene infusion bottles and the historical best monitoring accuracy requirements. If the deviation rate is less than or equal to the deviation limit, the initial identification baseline is used as the basis, and the fluctuation range is expanded based on the predicted particle size distribution value to form a particle size correction baseline, while the initial concentration identification baseline remains unchanged. Otherwise, based on the historical deviation correction data extracted from the multi-source monitoring historical database under the same working conditions, the fluctuation range of the initial concentration identification baseline is expanded, and the particle size correction baseline is recalculated by combining the predicted particle size distribution value and the historical deviation correction data. Finally, a corrected infusion foreign matter identification baseline containing particle size and concentration dimensions is generated.

[0122] An adaptive compensation algorithm is used to compensate for the two main sources of signal deviation in stages. This includes: for differences in material transmittance, a standard transmittance is obtained by combining the material transmittance requirements specified in the polypropylene infusion bottle production process standards with the material calibration values ​​of the production batch. A compensation coefficient is calculated based on the standard transmittance and the real-time estimated transmittance. The target amplitude is set at the standard signal amplitude range corresponding to the corrected infusion foreign body detection baseline. The amplitude of the original signal is then adjusted to the target amplitude range, effectively reducing signal weakening caused by material inhomogeneity, resulting in a compensated material deviation signal. The standard signal amplitude range is set based on the statistical values ​​of the signal amplitude from the historical best monitoring period. The original signal is acquired in real-time from the infusion sample to be monitored using an infusion foreign body detection device with calibrated core parameters.

[0123] For operational fluctuation deviations, the signal threshold range of the corrected infusion foreign body identification baseline is used as the screening criterion. Interference components that exceed the signal threshold range and match the operational fluctuation characteristics in the original signal are extracted. The signal phase is corrected to reduce the deviation between the signal phase and the standard phase of the corrected infusion foreign body identification baseline, thus obtaining an operational fluctuation deviation compensation signal that meets the requirements of the corrected infusion foreign body identification baseline. Among them, the operational fluctuation characteristics include, but are not limited to, the signal frequency corresponding to the uneven wall thickness of the polypropylene infusion bottle caused by bottle blowing pressure fluctuations.

[0124] Based on the weighted signal-to-noise ratio (SNR) fusion criterion, the material deviation compensation signal and the operating condition fluctuation deviation compensation signal are fused to form a full compensation signal. This full compensation signal is then compared feature-by-feature with the corrected infusion foreign body identification baseline. The online real-time identification result of the infusion foreign body is output. If the foreign body particle size in the full compensation signal is within the particle size threshold range of the corrected infusion foreign body identification baseline, the foreign body concentration is within the concentration threshold range of the corrected infusion foreign body identification baseline, and the SNR is greater than or equal to the industry acceptable threshold, the online identification result is considered normal. Otherwise, the online identification result is considered abnormal, including particle size exceeding the standard, concentration exceeding the standard, double exceeding the standard, false positives, and false negatives. The weighted signal-to-noise ratio fusion criterion is based on... Based on the reflectivity stability weight corresponding to the material type and the noise suppression weight corresponding to the operating flow rate, the signal-to-noise ratio of the material deviation compensation signal and the operating condition fluctuation deviation compensation signal are weighted and fused. The reflectivity stability weight is determined based on the matching degree between the material and the detection spectrum. For example, when the infusion tube is made of medical PVC material, the matching degree between PVC material and the detection spectrum is high, and the reflectivity stability weight is set to 0.55. The noise suppression weight is dynamically adjusted based on the standard deviation of the flow rate fluctuation. For example, when the standard deviation of the flow rate fluctuation is less than 0.2 m / s, the noise suppression weight is set to 0.45, which complements the reflectivity stability weight. The signal-to-noise ratio of the fused full compensation signal meets the real-time identification confidence requirements.

[0125] Specifically, the steps for determining the risk level of foreign body monitoring during intravenous infusion include:

[0126] According to the production batch, monitoring timestamp and operating condition of polypropylene infusion bottles, the online real-time identification results of foreign objects in the infusion are compared with the prediction results of the unsupervised association model to ensure that each set of data corresponds to the same infusion sample.

[0127] The foreign matter concentration deviation value is obtained by calculating the ratio of the absolute difference between the real-time measured foreign matter concentration value and the predicted foreign matter concentration value. If the predicted foreign matter concentration value is zero, the foreign matter concentration deviation value is calculated based on the real-time measured foreign matter concentration value and the maximum allowable concentration value for foreign matter detection in the polypropylene infusion bottle industry.

[0128] The intersection and union lengths of the real-time measured and predicted particle size distributions are determined by using the endpoint values ​​of the real-time measured and predicted particle size distribution intervals. The ratio of the intersection length to the union length is calculated to obtain the global overlap. The peak intersection and union lengths of the real-time measured and predicted particle size peak values ​​are determined by using the endpoint values ​​of the peak value intervals of the real-time measured and predicted particle size distributions. The peak value interval overlap is calculated. The peak value ratio deviation coefficient is determined by using the ratio of the real-time measured and predicted particle size peak values. The peak consistency is calculated by weighted summation based on the peak value interval overlap. Based on the peak consistency, the particle size matching degree is calculated by weighted summation based on the global overlap. The endpoint values ​​include the maximum and minimum values.

[0129] Based on the confidence level of real-time measured identification and the confidence level of foreign object type output by the unsupervised association model, the absolute difference of confidence level is calculated to determine the maximum confidence level between the real-time measured identification confidence level and the foreign object type confidence level. The confidence consistency index is obtained by subtracting the ratio of the absolute difference to the maximum confidence level from the base coefficient. In this embodiment, the base coefficient is set to a value of 1.

[0130] The concentration deviation threshold is set by optimizing the unsupervised correlation model based on the allowable range of industry concentration detection accuracy. The particle size matching threshold is set by the allowable range of particle size distribution detection error. The confidence consistency threshold is set according to the confidence reliability requirements.

[0131] A three-dimensional constraint rule is constructed, including a first constraint rule, a second constraint rule, and a third constraint rule. The first constraint rule requires that the concentration deviation value be less than or equal to the concentration deviation threshold, the particle size matching degree be greater than or equal to the particle size matching threshold, and the confidence consistency index be greater than or equal to the confidence consistency threshold. The second constraint rule requires that any two or any one of the concentration deviation value, particle size matching degree, and confidence consistency index meet the standard, such as the concentration deviation value being less than or equal to the concentration deviation threshold, the particle size matching degree being greater than or equal to the particle size matching threshold, and the confidence consistency index being less than the confidence consistency threshold. The third constraint rule requires that all three of the concentration deviation value, particle size matching degree, and confidence consistency index fail to meet the standard, such as the concentration deviation value being greater than the concentration deviation threshold, the particle size matching degree being less than the particle size matching threshold, and the confidence consistency index being less than the confidence consistency threshold.

[0132] Based on the three-dimensional constraint rules and combined with real-time anomaly identification results, the risk level of infusion foreign body monitoring is determined. When the first constraint rule is met and the infusion foreign body identification result is normal, it is determined to be a level 1 risk; when the second constraint rule is met and the particle size and concentration exceed the standard, it is determined to be a level 2 risk; when the third constraint rule is met and both exceed the standard, and there are false positives and false negatives, it is determined to be a level 3 risk. The risk increases progressively from level 1 to level 2 to level 3.

[0133] The tiered triggering mechanism does not send pause or warning instructions for Level 1 risks, and the batches of polypropylene infusion bottles at Level 1 risks continue to flow through the normal production process; for Level 2 risks, a warning instruction is sent to the production station control system but production is not suspended, and secondary testing is conducted using parameters that improve testing accuracy; for Level 3 risks, a pause instruction is immediately sent to the production station control system, and the batches of polypropylene infusion bottles at Level 3 risks are transferred to high-precision re-inspection equipment for both manual and equipment re-inspection.

[0134] After classifying risks at each level, samples of different risk levels are labeled according to type and entered into a multi-source monitoring dataset. For Level 2 and Level 3 risk samples, operating condition features and indicator deviation features are extracted and input into an unsupervised correlation model for incremental learning. Based on the reasons for deviations in Level 3 risk samples, the core parameters of the dynamic calibration mechanism are optimized, such as adjusting the baseline fluctuation range and exposure intensity calibration amplitude. Iterative optimization of the unsupervised correlation model and calibration mechanism is achieved through a risk sample-driven closed-loop learning mechanism, including using Level 2 and Level 3 risk samples as trigger samples and fine-tuning the parameters of the unsupervised correlation model through incremental learning strategies, such as fine-tuning graph convolution. The adjacency matrix weights of the network are adjusted, and the coefficients of the hidden layer neurons in the fully connected regression network are adjusted. New operating condition deviation patterns and foreign object feature patterns are learned. The parameters of the dynamic calibration mechanism are dynamically optimized according to the deviation reasons of the three-level risk samples. For example, if the deviation is caused by insufficient exposure intensity, the calibration range is increased from 5% to 8%. If the deviation is caused by flow rate fluctuation interference, the operating condition weight coefficients in the weighted signal-to-noise ratio fusion criterion are optimized. The optimized unsupervised association model and calibration mechanism are used for performance verification on the validation set. If the foreign object identification accuracy is lower than 85% and the prediction mean square error is greater than 0.05, it is judged as not meeting the standard; otherwise, it is judged as meeting the standard and saved as a new benchmark configuration to continuously reduce the false positive rate and false negative rate.

[0135] Example 2

[0136] Please see Figure 5 Another embodiment of the present invention provides: an online monitoring system for foreign bodies in infusion based on unsupervised self-learning, comprising: a multi-source fusion module, a correlation prediction module, a baseline dynamic calibration module, and a dual-source comparison and verification module;

[0137] The multi-source fusion module is used to collect multi-dimensional production parameter data and infusion foreign matter characteristic data of polypropylene infusion bottles under different production conditions in real time, and generate multi-source monitoring datasets after preprocessing.

[0138] The correlation prediction module is used to screen core working condition dimensions and determine the working condition stability weight coefficients based on the multi-source monitoring dataset. It combines the K-means clustering algorithm to divide the multi-source monitoring dataset. It extracts deep fusion features to construct an unsupervised correlation model, which learns and fits the dynamic correlation law between production parameters and infusion foreign body characteristics under each working condition and generates a correlation feature map. It inputs real-time production parameters and establishes a dual-branch collaborative output mechanism. The first branch outputs the infusion foreign body type and foreign body type confidence to determine the effectiveness of foreign body type identification. The second branch outputs the predicted value of infusion foreign body under the current working condition to determine the effectiveness of the predicted value and remove the predicted data corresponding to invalid infusion products.

[0139] The baseline dynamic calibration module is used to calculate the initial value of the identification baseline under each working condition by using the infusion foreign body characteristic data within the historical best monitoring period as a benchmark and combining the clustering centers of the correlation feature map under each working condition. Based on the effective infusion foreign body prediction value, a dynamic calibration mechanism for the infusion foreign body identification parameters is constructed to dynamically calibrate the core parameters of online monitoring of infusion foreign bodies. Based on the particle size distribution prediction value, a corrected infusion foreign body identification baseline is generated. Through an adaptive compensation algorithm, signal deviation is adaptively compensated, and the online real-time identification result of infusion foreign bodies is output.

[0140] The dual-source comparison module is used to compare the online real-time identification results with the prediction results of the unsupervised association model, calculate the three-dimensional indicators of the foreign body concentration deviation value, particle size matching degree and identification confidence consistency index, construct three-dimensional constraint rules in combination with preset thresholds, determine the risk level of infusion foreign body monitoring and trigger the graded processing mechanism, and iteratively optimize the unsupervised association model and calibration mechanism.

[0141] Working principle and effects:

[0142] By collecting multidimensional production parameter data and foreign matter characteristic data of polypropylene infusion bottles under different production conditions in real time, and generating multi-source monitoring datasets through preprocessing, efficient matching of production parameters and foreign matter characteristics at the working condition level is achieved, solving the problem of insufficient sample validity caused by data dispersion and lack of working condition correlation in traditional monitoring.

[0143] By selecting core operating condition dimensions and determining the operating condition stability weight coefficients, and combining the K-means clustering algorithm to divide the multi-source monitoring dataset, an unsupervised correlation model is constructed by extracting deep fusion features of production parameters and foreign body characteristics. The model learns and fits the dynamic correlation rules under each operating condition and generates a correlation feature map. After inputting real-time production parameters, the model outputs the infusion foreign body type, foreign body type confidence, and infusion foreign body prediction value through a dual-branch collaborative output mechanism. Invalid data is removed through validity judgment, which solves the problem of low infusion foreign body prediction accuracy caused by the inability of the unsupervised correlation model to adapt to dynamic changes in operating conditions.

[0144] Based on the infusion foreign body characteristic data within the historical best monitoring period, and combined with the clustering centers of the correlation feature map under various working conditions, the initial value of the identification baseline is calculated and an initial identification baseline is generated. A dynamic calibration mechanism is constructed based on the effective infusion foreign body prediction value to calibrate the core monitoring parameters. A corrected infusion foreign body identification baseline is generated based on the particle size distribution prediction value. At the same time, an adaptive compensation algorithm is used to compensate for the signal deviation caused by the difference in material transmittance and the fluctuation of working conditions, and outputs online real-time identification results. This solves the problem of high false positive and false negative rates of infusion foreign bodies caused by the traditional fixed baseline and uncompensated signal deviation.

[0145] By comparing the online real-time identification results with the prediction results of the unsupervised association model, three-dimensional indicators are calculated and three-dimensional constraint rules are constructed by combining them with preset thresholds. The risk level of foreign body monitoring in infusion is determined and a graded triggering mechanism is established. At the same time, the unsupervised association model and calibration mechanism are iteratively optimized, which solves the problem of accuracy decay caused by lack of verification of monitoring results and the inability of the model and mechanism to iterate autonomously.

[0146] Overall, a complete online monitoring system for foreign objects in infusions is formed through a four-layer architecture of multi-source fusion, correlation prediction, baseline dynamic calibration, and dual-source comparison. This system deeply integrates production conditions, parameter correlation, and foreign object identification, achieving high-precision identification and dynamic adaptation of foreign object characteristics in infusions. It effectively solves the problems of low identification accuracy, high false positive rate, and high false negative rate caused by dynamic changes in operating conditions, signal deviation, and insufficient model adaptation in foreign object monitoring of polypropylene infusion bottles, providing an efficient online monitoring solution for infusion production safety.

[0147] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for online monitoring of foreign bodies in intravenous infusions based on unsupervised self-learning, characterized in that, include: Acquire multi-source monitoring datasets, filter core operating condition dimensions and determine operating condition stability weight coefficients, and divide the multi-source monitoring datasets. An unsupervised association model is constructed to generate an association feature map. A two-branch collaborative output mechanism is established with real-time production parameters as input. The first branch outputs the type of infusion foreign body and the confidence level of the foreign body type to determine the identification effectiveness. The second branch outputs the predicted value of infusion foreign body under the current working conditions to determine the effectiveness of the predicted value. Calculate the initial value of the identification baseline under each working condition, generate the initial identification baseline, dynamically calibrate the core parameters of online monitoring of infusion foreign bodies based on the effective predicted value of infusion foreign bodies, generate the corrected infusion foreign body identification baseline, adaptively compensate for signal deviation, and output the online real-time identification results; The three-dimensional indicators are calculated and combined with preset thresholds to construct three-dimensional constraint rules, which are used to determine the risk level of foreign body monitoring in infusion and trigger a graded processing mechanism, and the unsupervised association model is iteratively optimized.

2. The online monitoring method for foreign bodies in intravenous infusion based on unsupervised self-learning according to claim 1, characterized in that, The specific steps for partitioning a multi-source monitoring dataset include: Based on the multi-source monitoring dataset, injection molding temperature, blow molding pressure and sterilization time were selected as the core operating condition dimensions. The analytic hierarchy process was used to determine the stability weight coefficients of the operating conditions, and the K-means clustering algorithm was used to cluster the core operating condition dimension data. The absolute difference between the real-time values ​​of each core operating condition parameter and the industrial benchmark value is calculated. Based on the ratio of the absolute difference to the industrial benchmark value, and combined with the operating condition stability weighting coefficient, the comprehensive parameter fluctuation amplitude is calculated. ; Set amplitude threshold , Divide the multi-source monitoring dataset; like Then it is divided into a subset of stable operating conditions; if Then it is divided into a subset of fluctuating operating conditions; if Then it is divided into a subset of abnormal operating conditions; Remove the abnormal operating condition subset and retain the stable operating condition subset and the fluctuating operating condition subset.

3. The online monitoring method for foreign bodies in intravenous infusion based on unsupervised self-learning according to claim 2, characterized in that, The specific steps for generating the associated feature map include: For the stable operating condition subset and the fluctuating operating condition subset, deep fusion features are extracted to generate a stable operating condition feature set and a fluctuating operating condition feature set; Construct an unsupervised association model, including an input layer, a feature fusion enhancement layer, a dual-task processing layer, and an output layer; The input layer receives the feature vectors of the stable operating condition feature set, the fluctuating operating condition feature set, and real-time production parameters; The feature fusion enhancement layer performs dimensional restructuring and correlation enhancement on the deep fusion features, and focuses on the strongly correlated features of production parameters and infusion foreign body characteristics through an attention mechanism to establish a correlation weight matrix; The dual-task processing layer utilizes a lightweight graph convolutional network in the type recognition branch to learn the mapping relationship between the fluctuation of working condition parameters and the type of foreign object and extract foreign object features based on the association weight matrix. The quantitative prediction branch uses a fully connected regression network to simultaneously learn the quantitative correlation between production parameters and foreign matter concentration and particle size distribution, and outputs a predicted feature vector.

4. The online monitoring method for foreign bodies in intravenous infusion based on unsupervised self-learning according to claim 3, characterized in that, The specific steps for generating the associated feature map also include: The output layer establishes a dual-branch collaborative output mechanism. The first branch outputs the type of foreign body in the infusion and the confidence level of the foreign body type, while the second branch outputs the predicted value of the foreign body in the infusion, including the predicted value of particle size distribution and the predicted value of foreign body concentration. The stable operating condition feature set and the fluctuating operating condition feature set are divided into a training set and a validation set according to a preset ratio; The unsupervised correlation model is trained based on the training set, using a joint loss function and the Adamw optimizer. After training, the performance of the unsupervised association model is verified using a validation set. The recognition accuracy and prediction mean square error of foreign object types are calculated, and accuracy thresholds and error thresholds are set. If the recognition accuracy is greater than or equal to the accuracy threshold and the prediction mean square error is less than or equal to the error threshold, the unsupervised association model is deemed to have met the standard. After meeting the standard, an association feature map is generated based on the association weight matrix. Otherwise, adjust the optimizer learning rate and retrain.

5. The online monitoring method for foreign bodies in intravenous infusion based on unsupervised self-learning according to claim 4, characterized in that, The specific steps for determining the validity of predicted values ​​include: The real-time production parameters are input into the unsupervised association model after training has reached the target. The first branch matches the type of foreign object and the confidence level of the foreign object type based on the association feature map. Set a confidence threshold to determine the effectiveness of foreign object type identification; If the confidence level of the foreign object type is greater than or equal to the confidence threshold, the foreign object type identification result is deemed valid; otherwise, a model warning is triggered and the current production parameter fluctuation information is recorded. The second branch outputs the predicted particle size distribution and the predicted foreign matter concentration under the current operating conditions; Set prediction thresholds, including a first prediction threshold and a second prediction threshold, to determine the validity of the predicted values; If the predicted particle size distribution is less than or equal to the first prediction threshold and the predicted foreign body concentration is less than or equal to the second prediction threshold, then the predicted foreign body concentration in the infusion is deemed valid. Otherwise, the predicted value of foreign matter in the infusion is deemed invalid, the corresponding infusion product is marked as unqualified, and the production line is activated to perform diversion processing.

6. The online monitoring method for foreign bodies in intravenous infusion based on unsupervised self-learning according to claim 5, characterized in that, The specific steps for generating the initial identification baseline include: Using the infusion foreign body characteristic data within a unified historical optimal monitoring period as a benchmark, cluster centers for each working condition are extracted from the associated feature map. The visualized association information of cluster centers is quantified into cluster feature vectors to form a cluster center feature set for each working condition; Calculate the correlation strength between the characteristics of each cluster center and the properties of infusion foreign bodies; If the correlation strength is greater than or equal to the preset screening threshold, the clustering feature is deemed valid and saved to the cluster center feature set of the corresponding working condition; otherwise, invalid clustering features are removed. For each working condition, calculate the mean of the best foreign object characteristics in history, extract the effective cluster center foreign object characteristic values, and calculate the first and second baseline initial values ​​of foreign object characteristics by weighting. Set integration weights, and integrate the initial values ​​of the first and second baselines to form the initial particle size identification baseline and the initial concentration identification baseline. Then, merge them according to the foreign matter characteristic dimension to generate the initial identification baseline.

7. The online monitoring method for foreign bodies in intravenous infusion based on unsupervised self-learning according to claim 6, characterized in that, The specific steps for generating a corrected baseline for foreign body identification in infusions include: Based on effective infusion foreign body prediction values, core parameters for online infusion foreign body monitoring are calibrated in multiple dimensions. In terms of exposure intensity, the near-infrared spectrum of the polypropylene infusion bottle is obtained, and the transmittance is estimated in real time based on the correspondence between the near-infrared spectrum and transmittance in order to calibrate the exposure intensity. In terms of signal acquisition frequency, adjustments are made in stages based on the amplitude of fluctuations in operating conditions. If the fluctuation range of the operating condition is less than or equal to the preset fluctuation threshold, the basic signal acquisition frequency is maintained; otherwise, the signal acquisition frequency is increased in stages. The pre-signal for foreign object detection is acquired in real time using the calibrated core parameters, and the deviation rate between the foreign object particle size distribution in the pre-signal and the predicted particle size distribution is calculated. If the deviation rate is less than or equal to the preset deviation limit, the fluctuation range is expanded based on the predicted particle size distribution to form a particle size correction baseline; Otherwise, based on historical deviation correction data under different operating conditions, the fluctuation range of the initial concentration identification baseline is expanded and the particle size correction baseline is recalculated; Finally, a corrected baseline for foreign body identification in infusion is generated.

8. The online monitoring method for foreign bodies in intravenous infusion based on unsupervised self-learning according to claim 7, characterized in that, The specific steps for adaptive compensation of signal bias include: Signal deviation is compensated step by step using an adaptive compensation algorithm; For differences in material transmittance, the standard transmittance is obtained and the real-time estimated transmittance is combined to calculate the compensation coefficient. The gain of the original signal amplitude is adjusted to obtain the material deviation compensation signal. For the deviation of operating conditions, the interference component that matches the characteristics of the operating conditions fluctuation is extracted from the original signal, and the deviation between the signal phase and the standard phase of the corrected infusion foreign body identification baseline is reduced to obtain the operating condition fluctuation deviation compensation signal. The material deviation compensation signal and the working condition fluctuation deviation compensation signal are fused to form a full compensation signal, which is then compared feature by feature with the corrected infusion foreign body identification baseline to output the online real-time identification result of infusion foreign body; If the particle size and concentration of foreign matter in the full compensation signal are within the threshold range corresponding to the corrected infusion foreign matter identification baseline and the signal-to-noise ratio is greater than or equal to the industry qualified threshold, then the online identification result is determined to be normal. Otherwise, the online recognition result will be deemed abnormal.

9. The online monitoring method for foreign bodies in intravenous infusion based on unsupervised self-learning according to claim 8, characterized in that, The specific steps for determining the risk level of foreign body monitoring during intravenous infusion include: By comparing the online real-time identification results with the prediction results of the unsupervised association model, the foreign matter concentration deviation value, particle size matching degree, and confidence consistency index are calculated. Set the concentration deviation threshold, particle size matching threshold, and confidence consistency threshold; Construct three-dimensional constraint rules and combine them with real-time online identification results to determine the risk level of foreign body monitoring in infusion. When the first constraint rule is met and the online recognition result is normal, it is judged as a level 1 risk; When the second constraint rule is met and both particle size and concentration exceed the standard, it is judged as a level 2 risk. When the third constraint rule is met and there are double exceedances, misjudgments, and missed detections, it is judged as a level three risk; After classifying and processing risks at each level, operating condition features and indicator deviation features are extracted for the secondary and tertiary risk samples and input into the unsupervised association model for incremental learning. The core parameters of dynamic calibration are optimized based on three levels of risk samples, and the unsupervised association model is iteratively optimized.

10. An online monitoring system for foreign bodies in intravenous infusions based on unsupervised self-learning, used to implement the online monitoring method for foreign bodies in intravenous infusions based on unsupervised self-learning as described in any one of claims 1-9, characterized in that, include: Multi-source fusion module, correlation prediction module, baseline dynamic calibration module, and dual-source comparison module; The multi-source fusion module is used to collect multi-dimensional production parameter data and infusion foreign matter characteristic data of polypropylene infusion bottles under different production conditions in real time, and generate multi-source monitoring datasets after preprocessing. The correlation prediction module is used to filter core working condition dimensions and determine the working condition stability weight coefficient, divide the multi-source monitoring dataset; construct an unsupervised correlation model, generate correlation feature map, input real-time production parameters, and establish a dual-branch collaborative output mechanism. The first branch outputs the infusion foreign body type and foreign body type confidence level to determine the identification effectiveness, and the second branch outputs the infusion foreign body prediction value under the current working condition to determine the prediction value effectiveness. The baseline dynamic calibration module is used to calculate the initial value of the identification baseline under each working condition, generate the initial identification baseline, dynamically calibrate the core parameters of online monitoring of infusion foreign bodies based on the effective predicted value of infusion foreign bodies, generate the corrected infusion foreign body identification baseline based on the predicted value of particle size distribution, adaptively compensate for signal deviation, and output the online real-time identification result of infusion foreign bodies. The dual-source comparison module is used to calculate three-dimensional indicators, construct three-dimensional constraint rules in combination with preset thresholds, determine the risk level of foreign body monitoring in infusion and trigger a graded processing mechanism, and iteratively optimize the unsupervised association model.

Images