A multi-parameter fusion endotoxin detection scheme optimization method and system

Abstract

The application discloses a multi-parameter fusion endotoxin detection scheme optimization method and system, relates to the technical field of detection, and comprises the following steps: loading input parameters of a to-be-detected sample; generating a prior detection scheme based on the input parameters and a pre-constructed multi-parameter knowledge graph; collecting multi-source signals in real time during detection according to the prior detection scheme; constructing a Bayesian state space model by taking the multi-source signals as input, taking temperature drift, edge evaporation effect and beta-glucan false positive interference as hidden variables for real-time estimation, and obtaining posterior endotoxin concentration estimation values and confidence intervals thereof; and taking the posterior endotoxin concentration estimation values as input, monitoring the satisfaction state of a hard constraint node specified in a pharmacopoeia in real time, and automatically triggering an online redilution instruction or a well site rearrangement instruction when it is predicted that any hard constraint node will not be satisfied. The application solves the problems of low detection pass rate and high uncertainty of detection results in the prior art.

Description

Technical Field

[0001] This invention relates to the field of detection technology, specifically to an optimization method and system for a multi-parameter fusion endotoxin detection scheme. Background Technology

[0002] Bacterial endotoxin testing is a critical step in the quality control of pharmaceuticals and medical devices. The accuracy of the test results directly affects product release decisions and patient safety.

[0003] Current methods mainly rely on the gelation method, dynamic turbidimetric method, and dynamic colorimetric method specified in the pharmacopoeia. The selection of detection schemes depends on human experience and trial and error, resulting in low detection pass rate and high uncertainty of detection results. Summary of the Invention

[0004] This application provides a method and system for optimizing endotoxin detection schemes through multi-parameter fusion, which addresses the problems of low detection pass rate and high uncertainty of detection results in existing technologies.

[0005] In view of the above problems, this application provides a method and system for optimizing endotoxin detection schemes through multi-parameter fusion.

[0006] Firstly, this application provides a method for optimizing a multi-parameter fusion endotoxin detection scheme, the method comprising: Load the input parameters of the sample to be tested, including sample type, estimated endotoxin concentration range, sample matrix information and historical interference records; Based on the input parameters and the pre-constructed multi-parameter knowledge graph, a priori detection scheme is generated. The priori detection scheme includes a recommended detection method, a recommended Limulus amebocyte lysate (LAL) reagent sensitivity, a recommended dilution gradient matrix, a recommended well plate layout scheme, and an estimated uncertainty value. The multi-parameter knowledge graph contains hard constraint nodes specified in the pharmacopoeia. During the detection process according to the prior detection scheme, multi-source signals are acquired in real time. The multi-source signals include dynamic curve data of each well position, real-time temperature data of each channel, well plate position information and sample volume monitoring data. Using the multi-source signals as input, a Bayesian state-space model is constructed. Temperature drift, edge volatilization effect and β-glucan false positive interference are used as latent variables for real-time estimation to obtain the estimated posterior endotoxin concentration and its confidence interval. Using the estimated endotoxin concentration as input, the system monitors in real time the satisfaction status of the hard constraint nodes specified in the pharmacopoeia, and automatically triggers an online redilution command or a well rearrangement command when it predicts that any hard constraint node will not be satisfied.

[0007] Secondly, the present invention provides a multi-parameter fusion endotoxin detection scheme optimization system, the system comprising: The input parameter acquisition module is used to load the input parameters of the sample to be tested. The input parameters include sample type, estimated endotoxin concentration range, sample matrix information and historical interference records. The test scheme generation module is used to generate a priori test scheme based on the input parameters and a pre-built multi-parameter knowledge graph. The priori test scheme includes a recommended test method, a recommended Limulus amebocyte lysate (LAL) reagent sensitivity, a recommended dilution gradient matrix, a recommended well plate layout scheme, and an estimated value of the expected uncertainty. The multi-parameter knowledge graph contains hard constraint nodes specified in the pharmacopoeia. The multi-source signal acquisition module is used to acquire multi-source signals in real time during the detection process according to the prior detection scheme. The multi-source signals include dynamic curve data of each well position, real-time temperature data of each channel, well plate position information and sample volume monitoring data. The posterior estimation module is used to construct a Bayesian state-space model with the multi-source signal as input, and to estimate the temperature drift, edge volatilization effect and β-glucan false positive interference as latent variables in real time to obtain the posterior endotoxin concentration estimate and its confidence interval. The instruction triggering module is used to monitor the satisfaction status of the hard constraint nodes specified in the pharmacopoeia in real time, using the estimated posterior endotoxin concentration as input, and automatically trigger an online redilution instruction or a well rearrangement instruction when it is predicted that any hard constraint node will not be satisfied.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application first loads the input parameters of the sample to be tested, providing basic data for the prior distribution. Second, based on the input parameters and a multi-parameter knowledge graph, a prior detection scheme is generated as initial data for optimization. Through prior guidance, the intelligent generation of the prior detection scheme is achieved. Furthermore, during the detection process of the prior detection scheme, multi-source signals are collected in real time, providing basic data for subsequent updates and optimizations of the detection scheme. Further, using the multi-source signals as input, a Bayesian state-space model is constructed. Through real-time estimation of latent variables, the estimated posterior endotoxin concentration and its confidence interval are obtained, and the posterior distribution is acquired as the discrimination criterion for subsequent updates, improving detection efficiency and shortening detection time. Finally, using the estimated posterior endotoxin concentration as input, the satisfaction status of hard constraint nodes specified in the pharmacopoeia is monitored. By predicting the satisfaction status, it is determined whether to issue an instruction, improving the reliability of the detection, shortening the detection time, and triggering an online redilution instruction or well rearrangement instruction when the prediction result indicates a non-satisfaction status. Virtual playback of the digital twin model is performed, and the knowledge graph is adaptively updated through verification of effective detection results, achieving real-time prediction and automatic intervention for compliance checks. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating an optimization method for a multi-parameter fusion endotoxin detection scheme provided in this application; Figure 2 This is a schematic diagram of the structure of an optimized system for multi-parameter fusion endotoxin detection scheme provided in this application.

[0010] In the attached diagram, the components represented by each number are as follows: Input parameter acquisition module 11, verification scheme generation module 12, multi-source signal acquisition module 13, posterior estimation module 14, and instruction triggering module 15. Detailed Implementation

[0011] This application provides an optimized method and system for multi-parameter fusion endotoxin detection, which specifically addresses the problems of low detection pass rate and high uncertainty of detection results in the prior art.

[0012] The present invention will now be described in detail with reference to the accompanying drawings.

[0013] Example 1, as Figure 1 As shown, this application provides a method for optimizing a multi-parameter fusion endotoxin detection scheme, the method comprising: S10: Load the input parameters of the sample to be tested, including sample type, estimated endotoxin concentration range, sample matrix information and historical interference records; In this embodiment, the sample type is a classification of the nature of the sample to be tested, such as active pharmaceutical ingredient, injection, medical device, biological product, traditional Chinese medicine injection, etc.; the estimated endotoxin concentration range is the range of endotoxin content estimated based on historical data of the sample or information of the same type; the sample matrix information is a description of the composition of the sample, including whether it contains proteins, sugars, lipids, surfactants, chelating agents, pH deviations, and other substances that may interfere with the detection; historical interference records are abnormal records that have occurred in the past detection of the sample or similar samples, such as unqualified recovery rate, false positive, false negative, Limulus amebocyte lysate (LAL) reagent mismatch, exceeding the standard curve range, etc.

[0014] Specifically, the system retrieves the sample to be tested from the background database, including input parameters such as sample type, estimated endotoxin concentration range, sample matrix information, and historical interference records. These parameters serve as input parameters for subsequent multi-parameter knowledge graph generation, providing a data foundation for the process.

[0015] S20: Based on the input parameters and the pre-constructed multi-parameter knowledge graph, generate a priori detection scheme, which includes a recommended detection method, a recommended Limulus amebocyte lysate (LAL) reagent sensitivity, a recommended dilution gradient matrix, a recommended well plate layout scheme, and an estimated uncertainty value, wherein the multi-parameter knowledge graph contains hard constraint nodes specified in the pharmacopoeia; In the embodiments of this application, the recommended detection method can recommend feasible or applicable detection methods for the sample to be tested, such as dynamic colorimetric method; the recommended Limulus Amebocyte Lysate (LAL) reagent sensitivity is matched with the expected endotoxin concentration, such as LAL reagent batch number and labeled sensitivity, such as 0.03 EU / mL; the recommended dilution gradient matrix is ​​a recommended combination of sample dilution factors used to construct a standard curve or perform sample determination.

[0016] The recommended well plate layout scheme is a scheme for allocating the specific well coordinates and the number of replicate wells for standards, test samples, positive controls, and negative controls on the microplate; the expected uncertainty estimate is the uncertainty of the final test result based on the pass rate of historical cases and the interference risk assessment of the current scheme.

[0017] Specifically, based on input parameters and a multi-parameter knowledge graph, constraints are imposed according to hard constraint nodes specified in the pharmacopoeia, and the recommended detection method, recommended Limulus amebocyte lysate (LAL) reagent sensitivity, recommended dilution gradient matrix, recommended well plate layout scheme, and expected uncertainty estimate for the sample to be tested are determined accordingly. Then, a priori detection scheme for the sample to be tested is constructed.

[0018] In step S20 of the method provided in this application embodiment, before loading the input parameters of the sample to be tested, the method further includes the step of constructing the multi-parameter knowledge graph: Obtain the hard constraint rules specified in the pharmacopoeia and convert them into Class I nodes. The hard constraint rules include at least the maximum effective dilution factor constraint, the sensitivity qualified interval constraint, the recovery rate range specified in the pharmacopoeia constraint, the correlation coefficient of the standard curve threshold constraint specified in the pharmacopoeia constraint, and the pH value after mixing range specified in the pharmacopoeia constraint. The soft constraint attributes in historical test records are obtained and transformed into second-type nodes. The soft constraint attributes include at least sample type, batch-to-batch variation coefficient of Limulus amebocyte lysate reagent, applicable scenario label of test method, and historical deviation data of instrument channel. Establish the association between the first type of nodes and the second type of nodes to form a multi-layered knowledge graph structure.

[0019] In this embodiment, firstly, the hard constraint rules specified in the pharmacopoeia for bacterial endotoxin testing are obtained by analyzing the relevant provisions. Then, these hard constraint rules are transformed into first-type hard constraint nodes. Among the hard constraint rules, the maximum effective dilution factor constraint is the maximum permissible dilution factor specified in the pharmacopoeia for a specific sample type, such as water for injection or biological products. Exceeding this factor will result in excessive dilution of the endotoxin concentration, falling below the detection limit of the Limulus Amebocyte Lysate (LAL) reagent, leading to false negative results.

[0020] The sensitivity acceptable range constraint is the allowable deviation range between the calibrated value and the labeled value of the sensitivity of the horseshoe crab reagent as specified in the pharmacopoeia. Only within this range is the batch of horseshoe crab reagent considered acceptable. The recovery rate range constraint specified in the pharmacopoeia is the ratio of the measured value to the theoretical value after adding standard endotoxin to the sample. The pharmacopoeia usually specifies that it should be between 50% and 200%, which is used to evaluate the interference or enhancement effect of the sample matrix on the detection.

[0021] The pharmacopoeia-specified threshold constraint for the correlation coefficient of the standard curve means that in quantitative detection, the linear correlation coefficient between the logarithmic value of the endotoxin concentration and the reaction time must be greater than or equal to the threshold specified in the pharmacopoeia; otherwise, the standard curve is invalid. The pharmacopoeia-specified range constraint for the pH value after mixing means that the pH value of the mixture after the sample to be tested is mixed with the Limulus amebocyte lysate (LAL) reagent must fall within the range specified in the pharmacopoeia, usually 6.0-8.0.

[0022] Secondly, soft constraint attributes from historical detection records are obtained and transformed into second-type nodes. Soft constraint attributes, unlike hard constraints, are non-enforceable but can significantly affect the success rate, accuracy, or efficiency of the detection scheme. Violating soft constraints usually does not directly lead to invalid detection, but it increases the risk of detection failure or reduces the confidence of the results.

[0023] The soft constraint attributes include at least sample type, batch-to-batch coefficient of variation of the horseshoe crab reagent, applicable scenario label for the detection method, and historical deviation data of the instrument channels. Sample type refers to the category of the sample to be tested, such as water for injection. The batch-to-batch coefficient of variation of the horseshoe crab reagent is the degree of difference in detection sensitivity and reaction kinetic parameters between different production batches of the reagent. It is obtained through statistical analysis after testing multiple batches of reagents with standards and is a key quantitative indicator for evaluating the batch consistency of the horseshoe crab reagent.

[0024] The applicable scenario label for the detection method is based on historical successful cases, indicating the optimal performance of a specific detection method under what sample type, concentration range, and interference level. The historical deviation data for instrument channels is a statistical analysis of the deviations exhibited by different well positions of a multi-channel detection instrument relative to a reference channel over long-term use.

[0025] After obtaining the soft constraint attributes, they are transformed into the second type of nodes in the knowledge graph structure, and probabilistic risk warnings are obtained through the second type of nodes.

[0026] Finally, the first type of hard-constrained nodes and the second type of soft-constrained nodes are connected by associated edges to establish the logical relationship between the hard-constrained nodes and the soft-constrained nodes, and the influence relationship between the nodes, such as weight and relationship label, is determined. Finally, the constructed nodes and edges are stored to form a hierarchical network structure that supports traversal query and reasoning calculation, namely a multi-layer knowledge graph structure.

[0027] Step S20 in the method provided in this application embodiment includes: The input parameters are encoded into query vectors, and the multi-parameter knowledge graph is called to perform path retrieval. Historical cases that have a similarity to the input parameters that meet a preset standard are matched as matching results. Based on the matching results, the optimal solution is searched in the preset candidate space of solutions using a Bayesian optimization model. The detection pass rate of the historical cases is used as a priori, and the prior detection solution is output after iterative evaluation.

[0028] In this embodiment, firstly, an encoder maps the discrete and continuous features of the parameters to a query vector. The query vector is then input, and a multi-parameter knowledge graph is invoked for path retrieval. The cosine similarity between the query vector and each historical case vector stored in the graph is calculated. The cosine similarity is the ratio of the dot product of the query vector and the product of the modulo lengths of the historical case vectors. Historical cases whose similarity to the input parameters reaches a preset standard are considered matching results. The preset standard is a pre-defined criterion; only historical cases with a similarity to the query vector not lower than the preset standard are considered to have sufficient reference value and are included in the matching results. This standard can be obtained based on the average similarity between historical case vectors, for example, a cosine similarity ≥ 0.85.

[0029] Secondly, based on the matching results, the optimal solution is searched in the pre-defined candidate space of the solution by the Bayesian optimization model. The detection pass rate of historical cases is used as a priori. After iterative evaluation, the prior detection solution is output. The pre-defined candidate space of the solution is a set of feasible solutions predefined in the process of generating the prior solution. It consists of all possible combinations of decision variables such as detection method, sensitivity specification of Limulus amebocyte lysate reagent, dilution gradient factor, and well plate layout, which are used by the Bayesian optimization model to search for the optimal solution.

[0030] Specifically, based on the matching results, during the initialization of the Bayesian optimization model, the matching results are used as initial evaluation points in a predefined candidate space based on existing historical solutions. The pass rate of the matched historical cases is used as the benchmark for the Bayesian optimization surrogate model to initially predict the pass rate of unknown candidate solutions. Internally, the pass rate of all historical cases in the matching results is calculated and used as the hyperparameter of the Gaussian process's mean function at the prior point. Subsequently, the pass rate of all historical cases is calculated iteratively multiple times, stopping after a predefined number of iterations, and the prior detection solution is output.

[0031] In this embodiment, the input parameters are encoded into query vectors, and path retrieval and similarity matching are performed in a multi-parameter knowledge graph to obtain the most similar historical successful cases for the current sample. The detection pass rate of matching historical cases is used as the prior distribution for Bayesian optimization. The optimal solution output after iterative evaluation has a much higher expected pass rate than random selection or fixed procedures, reducing invalid detections and improving the success rate of the detection solution. Through the Bayesian optimization model, the expected uncertainty estimate of the solution detection result is given based on the posterior probability distribution, realizing efficient search in high-dimensional parameter space.

[0032] S30: During the detection process according to the prior detection scheme, multi-source signals are collected in real time. The multi-source signals include dynamic curve data of each well position, real-time temperature data of each channel, well plate position information and sample volume monitoring data. In this embodiment, the multi-source signals are various data types reflecting the experimental process, acquired from different physical or logical sources within the detection system. The kinetic curve data for each well is a sequence curve showing the change in optical density or turbidity value of each well in the microplate over time, continuously measured by the microplate reader during incubation. This curve directly characterizes the reaction kinetics of endotoxin and Limulus amebocyte lysate (LAL) reagent. The real-time temperature data for each channel is a real-time temperature sequence collected by temperature sensors located at different positions within the incubation chamber.

[0033] The well plate position information is the row and column coordinates of each well in the microplate, which is associated with the kinetic curve data and used to identify edge effects and position-dependent systematic biases; the sample volume monitoring data is the real-time volume feedback value generated by the autosampler each time liquid is aspirated or discharged, which is used to detect whether there is insufficient or excessive sample addition.

[0034] Specifically, after the detection process begins, a multi-source signal real-time acquisition step is initiated. A synchronized data stream is established with the microplate reader, temperature sensor array, and autosampler via an industrial Ethernet or USB interface. Taking a time step Δt = 20 seconds as an example, at each sampling moment, the optical density values ​​at 405nm wavelength for all occupied wells are read from the microplate reader, forming a kinetic curve data array that grows over time. Simultaneously, the real-time temperature values ​​of each well are read from the chamber temperature sensors, forming a temperature distribution vector; the actual sample volume allocated to each well is read from the autosampler's injection pump encoder, forming a volume deviation record; and the coordinates of the plate position corresponding to each kinetic data point are associated. All data are appended with a global timestamp at the moment of acquisition and stored in a circular buffer.

[0035] In this embodiment of the application, during the execution of the prior detection scheme, multi-source signals are collected in real time to realize the perception of the detection process and provide synchronous process observation data for the subsequent Bayesian state-space model.

[0036] S40: Using the multi-source signal as input, construct a Bayesian state-space model, and use temperature drift, edge volatilization effect and β-glucan false positive interference as latent variables for real-time estimation to obtain the estimated value of posterior endotoxin concentration and its confidence interval. In this embodiment, the Bayesian state-space model is a dynamic model based on Bayesian statistical inference, consisting of a state equation and an observation equation. Using Bayes' theorem, the probability distribution estimate of latent variables is continuously updated after obtaining new observation data. Latent variables are latent variables that cannot be directly measured but have a significant impact on the observation data. Temperature drift refers to the slow change and spatial non-uniformity of the actual temperature of the chamber relative to the set temperature during incubation. Temperature drift directly affects the rate constant of the enzymatic reaction of the horseshoe crab reagent. The edge evaporation effect occurs when uneven temperature distribution or prolonged exposure time at the edge wells of the microplate causes evaporation of the reaction solution and a reduction in volume, leading to deviations of the detection signal from the true value.

[0037] β-glucan false positive interference occurs when β-glucan present in the sample non-specifically activates the G factor pathway of the Limulus Amebocyte Lysate (LAL) reagent, generating a reaction signal similar to endotoxin. Real-time estimation involves updating the posterior distribution of latent variables as new observation data blocks are obtained over time. The posterior endotoxin concentration estimate and its confidence interval are calculated using Bayesian inference to determine the true probability distribution of endotoxin concentration after fusing prior knowledge and all real-time observation data.

[0038] Specifically, a Bayesian statistical inference dynamic model is constructed by using multi-source signals as input to the model. Then, temperature drift, edge volatilization effect, and β-glucan false positive interference are used as latent variables to estimate the significant impact of latent variables on the input data, and to infer the calculated probability distribution of the true concentration of endotoxin and the confidence interval.

[0039] Step S40 in the method provided in this application embodiment includes: Define a state vector, which includes the state variables of endotoxin true concentration, temperature deviation, dilution error, β-glucan interference intensity, and edge volatility effect. Establish an observation equation, and express the measured dynamic curve signal as a function of the state variable and the sum of the observation noise; The particle filter algorithm is used to perform Bayesian inference on the state space model. The posterior distribution is approximated through state transition and weight resampling iteration, and the contribution of each latent variable is separated in real time. Based on the Bayesian inference results, the posterior endotoxin concentration estimate and the confidence interval of the estimate are output.

[0040] In this embodiment, firstly, the state vector is a multidimensional array containing several state variables in the state space model; the endotoxin true concentration state variable is the true concentration of endotoxin in the sample to be tested; and the temperature deviation state variable is the deviation value of the actual temperature of a specified well in the incubation chamber from the set temperature.

[0041] The dilution error state variable is the logarithmic deviation of the actual concentration from the theoretical concentration during sample addition or dilution operations; the β-glucan interference intensity state variable is the equivalent endotoxin activity concentration produced by β-glucan present in the sample; the edge volatility effect state variable is the factor that increases the relative concentration of endotoxins and reagents in the wells due to the evaporation of liquid in the edge wells of the microplate.

[0042] Specifically, when constructing a Bayesian state-space model, the state vector is first defined. Based on prior knowledge and physicochemical principles, the state vector is set. Specifically, for a specific well where endotoxin detection is currently being performed, the actual endotoxin concentration corresponding to the well is used as the state variable to be estimated, and its initial value is set by the predicted concentration range in the prior detection scheme.

[0043] For example, the initial value of the temperature deviation state variable is set to 0, and the state transition equation at each time step is then allowed to perform a random walk with a small variance to fit the readings of the actual temperature sensor; the initial value of the dilution error state variable is set to 0, and the possible cumulative error is described by a first-order autoregressive model.

[0044] The initial value of the β-glucan interference intensity was set based on the historical interference records of this sample type in the knowledge graph and allowed to drift slowly over time; the initial value of the edge volatilization effect was 1, and an exponential growth model was used to simulate the accelerated evaporation process.

[0045] Secondly, the kinetic curve signal is a curve generated by continuously monitoring the reaction parameters of each well site over time in the dynamic detection method. For the turbidimetric method, the kinetic curve signal is the curve of transmittance or absorbance changing with time; for the colorimetric method, the kinetic curve signal is the curve of absorbance changing with time. The kinetic curve reflects the formation rate of the reaction product between endotoxin and Limulus amebocyte lysate (LAL) reagent.

[0046] Specifically, a function of state variables is constructed based on the biochemical kinetics principle of the Limulus amebocyte lysate (LAL) reagent reaction. Then, the sum of observation noise is obtained through cumulative calculation to obtain the kinetic curve signal. The constructed observation equation is a function of state variables and the sum of observation noise.

[0047] Furthermore, the particle filtering algorithm is used to perform Bayesian inference on the state-space model. The particle filtering algorithm is a Bayesian inference technique for non-Gaussian state-space models. It approximates the posterior probability distribution of the state variables through a set of weighted random samples. The particles evolve recursively through three steps: prediction, weight update, and resampling. Bayesian inference is the process of calculating the posterior distribution by using Bayes' theorem, combined with the prior distribution and the likelihood function.

[0048] Bayesian inference is performed through particle prediction, weight updates, and resampling. Each particle from the previous time step is transitioned to the predicted particle at the current time step via state transition. After obtaining the predicted observation data, the probability of each predicted particle's appearance is calculated, and its weight is updated. Then, particles are resampled based on the updated weight distribution; particles with high weights are replicated, while those with low weights are discarded, generating a next-generation particle set with equal weights. After resampling, the particle set again approximates the posterior distribution. This iterative process is repeated, gradually approximating the true posterior probability density function of the particles' empirical distribution. Finally, the marginal distributions of each state variable are extracted from the particle's posterior distribution, allowing for real-time separation of the contribution of each latent variable to the observed signal.

[0049] For example, suppose the total number of particles is N=5000. When initializing t=0, 5000 particles are generated according to the prior distribution. The initial weight of all particles is set to 1 / N. The posterior estimate of the endotoxin concentration is the weighted average of all particles, and its 95% confidence interval is the posterior 2.5% and 97.5% quantiles of the endotoxin concentration.

[0050] Finally, after completing the particle filtering update, the Bayesian inference results are obtained. The endotoxin concentration of each particle in the inference results is extracted, and the posterior mean and posterior standard deviation are calculated. Then, the posterior distribution is approximated as a Gaussian distribution to obtain the confidence interval. The mean endotoxin concentration of the most stable posterior estimation stage among all time points is taken as the final estimate, and the posterior endotoxin concentration estimate and its confidence interval are output.

[0051] In step S40 of the method provided in this application embodiment, the prior distribution of each latent variable in the Bayesian state space model is set based on the historical fluctuation range and fluctuation frequency recorded by the corresponding node in the multi-parameter knowledge graph.

[0052] In this embodiment, the historical fluctuation range is the range of values ​​in which the latent variable recorded by the node has appeared in a large number of past tests. For example, the historical fluctuation range of the temperature deviation of a certain model of ELISA reader is [-0.3℃, +0.2℃]. The historical fluctuation frequency is the number of records on which the historical fluctuation range is based, that is, the number of times the latent variable has been observed or calculated in the past. The higher the frequency, the higher the reliability of the statistical information.

[0053] Specifically, in the initialization phase of the Bayesian state-space model construction, for each latent variable to be defined, a multi-parameter knowledge graph is accessed. Historical records matching the current detection conditions are extracted from the multi-parameter knowledge graph, and the historical fluctuation amplitude and frequency of temperature drift, edge volatility effect, and β-glucan false positive interference are statistically analyzed. These are used as the mean and variance parameters of the corresponding latent variable prior distribution, thereby achieving the quantitative setting of the prior distribution.

[0054] In this embodiment, by defining a state vector containing multiple state variables, an observation equation is established that represents the dynamic curve signal as a function of the state variables. This enables Bayesian inference to infer the temporal evolution of each latent variable, achieving accurate separation of multi-source interference. The particle filtering algorithm approximates the state space with a large number of particles, handling nonlinear and non-Gaussian state transitions and observation models, as well as the contribution of real-time interference, thus improving the accuracy of endotoxin concentration estimation. The Bayesian inference results directly provide the posterior probability distribution, extracting intervals at arbitrary confidence levels, providing reliable statistical basis for hard constraint monitoring and compliance judgment.

[0055] S50: Using the estimated posterior endotoxin concentration as input, monitor the satisfaction status of the hard constraint nodes specified in the pharmacopoeia in real time, and automatically trigger an online redilution command or well rearrangement command when it is predicted that any hard constraint node will not be satisfied.

[0056] In this embodiment, the hard constraint node is a mandatory constraint node derived from pharmacopoeia rules; the satisfaction state is whether each hard constraint has been satisfied, whether it is currently satisfied, and whether it will be satisfied in the future, based on the currently obtained detection data and posterior estimation results; the prediction that any hard constraint node will not be satisfied is based on the current posterior estimation and trend, and by extrapolation or simulation, it is determined that the probability of the hard constraint being violated exceeds the set risk threshold before the end of the detection.

[0057] The online redilution command is an automatically issued operation command that requires the fully automated detection system to further dilute the current sample to be tested; the well rearrangement command is an operation command that moves the sample from certain wells to other wells.

[0058] Specifically, the inferred posterior endotoxin concentration estimate is used as input. While the Bayesian state-space model continuously outputs the posterior estimate and its confidence interval, a hard constraint monitor runs in parallel. The monitor takes the posterior estimate at each time step as input and checks whether the current state is satisfied based on the numerical boundaries of the hard constraint nodes in the multi-parameter knowledge graph. The state-space model is then used to predict the state trajectory and its confidence interval for several future time steps, and the probability of each hard constraint being violated is assessed.

[0059] When the predicted hard constraint node is not satisfied, an intervention command is generated, triggering an online redilution command or a well repositioning command to move the test specimen well from the edge position to an empty well in the central region. Finally, the newly collected data will be re-entered into the state-space model to form a closed-loop adaptive control.

[0060] Step S50 in the method provided in this application embodiment includes: A multi-objective optimization problem is constructed, with the objective function being to minimize the total uncertainty. The decision variables include the dilution factor, the sensitivity of the horseshoe crab reagent, the sample volume, and the incubation time. The constraints include all hard constraint nodes. The second generation of the non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem and generate a Pareto optimal front solution set. Based on the Pareto optimal frontier, the optimal solution that satisfies the current detection conditions is selected. The optimal solution includes the optimized values ​​for dilution factor, Limulus amebocyte lysate (LAL) reagent sensitivity, sample loading volume, and incubation time. The optimal dilution factor in the optimal solution is compared with the current dilution factor. If the optimal dilution factor is greater than the current dilution factor, an online re-dilution command is triggered to supplement the dilution of the sample according to the optimal dilution factor. The well plate layout parameters in the optimal solution are compared with the current well position. If the correlation coefficient of the standard curve is predicted to be lower than the threshold specified in the pharmacopoeia, a well position rearrangement instruction is triggered to move the sample to the target well position corresponding to the well plate layout parameters in the Pareto optimal frontier.

[0061] In this embodiment, a multi-objective optimization problem is first constructed, with the objective function being to minimize the total uncertainty. Decision variables include dilution factor, horseshoe crab reagent sensitivity, sample loading volume, and incubation time. Constraints include all hard constraint nodes. Specifically, the dilution factor is the pretreatment dilution ratio of the sample; the horseshoe crab reagent sensitivity is the nominal sensitivity of the selected reagent batch; the sample loading volume is the volume of sample and reagent added to each well; the incubation time is the reaction incubation duration; and the constraints include all hard constraint nodes specified in the pharmacopoeia, such as the maximum effective dilution factor, the acceptable sensitivity range, the pharmacopoeia-specified recovery rate range, the pharmacopoeia-specified threshold for the correlation coefficient of the standard curve, and the pharmacopoeia-specified range for the pH value after mixing.

[0062] Minimizing the total uncertainty comprehensively reflects the dispersion of the detection results and the width of the confidence interval. By solving the objective function to minimize the total uncertainty, a Pareto optimal solution set that satisfies all constraints is generated.

[0063] Secondly, the second generation of the non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem and generate the Pareto optimal front solution set. The second generation of the non-dominated sorting genetic algorithm is a widely used multi-objective evolutionary algorithm that can efficiently solve optimization problems with multiple conflicting objective functions.

[0064] Specifically, for each individual, the objective function value is calculated by calling the observation equation of the Bayesian state-space model and the current latent variable estimates, and it is checked whether all hard constraints are satisfied. Individuals that do not meet the constraints are assigned an infinite objective value, and iterative non-dominated sorting genetics is performed.

[0065] In the non-dominated sorting genetic iteration process, the population is first divided into different levels of frontiers based on the dominance relationships between solutions. For example, solution A dominates solution B if and only if A is no worse than B on all objectives and better than B on at least one objective. Then, crowding distance is calculated, and the diversity of solutions within the same frontier is evaluated to prioritize the retention of evenly distributed solutions. Through tournament selection, crossover, and mutation, the biological evolution process is simulated to generate a new generation of solutions. Finally, through an elite retention strategy, the parent and offspring generations are merged, and the optimal N individuals are selected to enter the next generation. After the cycle ends, all individuals in the final non-dominated layer 1 of the population are taken as the Pareto optimal frontier solution set.

[0066] Next, based on the Pareto optimal frontier solution set, according to the pre-set decision rules or the current detection priority, the optimal solution that meets the current detection conditions is selected to obtain the most suitable solution for actual intervention. Subsequently, for each selected solution, the optimized values ​​for dilution factor, Limulus amebocyte lysate (LAL) reagent sensitivity, sample loading volume, and incubation time are calculated, compared, and the optimal solution is output. Among them, the optimized dilution factor is the recommended total dilution factor for the sample in the selected solution; the optimized Limulus amebocyte lysate (LAL) reagent sensitivity is the recommended labeled sensitivity of the Limulus amebocyte lysate in the selected solution; the optimized sample loading volume is the recommended sample or reagent loading volume per well in the selected solution; and the optimized incubation time is the recommended remaining incubation or total incubation time in the selected solution.

[0067] Furthermore, the optimized dilution factor in the optimal solution is compared with the current dilution factor. If the optimized dilution factor is greater than the current dilution factor, the trigger condition is met, an online re-dilution instruction is generated, and the sample is further diluted according to the optimized dilution factor. For example, the optimized dilution factor in the optimal solution is, for instance, 8 times, and the actual dilution factor of the current sample is, for instance, 4 times. Since 8 > 4, the trigger condition is met, and a re-dilution is performed with the optimized dilution factor of 8 times.

[0068] Finally, the well plate layout parameters in the optimal solution are compared with the current well positions, and the real-time predicted value of the correlation coefficient of the standard curve is obtained from the current detection state. If the current prediction shows that the real-time predicted value of the correlation coefficient of the standard curve is lower than the pharmacopoeia threshold, well position rearrangement is triggered, and the sample is moved to the target well position corresponding to the well plate layout parameters in the Pareto optimal front for well position rearrangement.

[0069] The method provided in this application embodiment also includes: During the detection process, the entire process data is continuously recorded. The entire process data includes real-time temperature sequences at each time point, dynamic curve data of each well, timestamps of each step, sample volume data, equipment parameter settings, and state vector estimates at each time point. After the test is completed, a digital twin model is constructed based on the full-process data and a virtual replay is performed. If the test result is out of standard, the root cause parameter causing the out-of-standard result is located through the virtual replay.

[0070] In this embodiment of the application, firstly, during the time from the start to the end of the detection process, the entire process data is recorded at a fixed sampling rate. The entire process data includes the real-time temperature sequence at each time point, the dynamic curve data of each well, the timestamp of each step of the operation, the sample volume data, the equipment parameter settings, and the state vector estimation value at each moment.

[0071] Among them, the real-time temperature sequence at each time point is the temperature value recorded at multiple locations in the incubation chamber at fixed time intervals, forming a two-dimensional time-temperature array; the kinetic curve data of each well is the optical density or turbidity value measured at fixed time intervals during the detection process of each microwell, forming an array; the timestamp of each operation step is the time point of each operation process of the device, including the start / end time of sample addition, the start or end time of mixing, the start time of measurement by the microplate reader, etc.

[0072] The sample addition volume data is the actual dispensing volume fed back by the flow sensor or syringe pump encoder after each sample addition step; the equipment parameter settings are the configuration parameters of each device during the detection process, including the measurement wavelength, gain value, and vibration plate parameters of the microplate reader.

[0073] The state vector estimates at each time step are the posterior state vector estimates output by the Bayesian state-space model at each sampling time step, including endotoxin concentration estimates, temperature deviation estimates, dilution error estimates, β-glucan interference intensity estimates, edge volatility factor estimates, and their confidence intervals.

[0074] Secondly, after the test is completed, a digital twin model of the out-of-specification detection task is constructed based on the recorded full-process data, and virtual playback is performed to dynamically replay the entire detection process in the digital twin model. If the test result shows that the endotoxin concentration exceeds the acceptable limit specified by the pharmacopoeia or the company, the root cause parameter causing the out-of-specification can be located through virtual playback. After adjusting the root cause parameter, the out-of-specification result can be reproduced or eliminated.

[0075] In step S50 of the method provided in this application embodiment, a digital twin model is constructed and virtual playback is performed. If the detection result is out of standard, the root cause parameter causing the out-of-standard is located through the virtual playback, including: A timeline for the corresponding detection task is constructed based on the full-process data. Based on the timeline and the full-process data, a digital twin model is generated, which includes physical level information, reaction dynamics level information, state estimation level information and decision level information. The digital twin model is presented virtually through an interactive interface; If the test result is out of standard, the deviation point is found by comparing the digital twin model of the normal sample that was successfully tested in the same batch, and the earliest occurrence position corresponding to the deviation point is deduced from the detection time axis. Then, the parameters of the earliest occurrence position are compared with the normal range item by item to determine one or more root cause parameters.

[0076] In this embodiment, firstly, a timeline for the corresponding detection task is constructed based on the full-process data, where the timeline is a continuous linear coordinate system with time as the independent variable. Then, timestamps of all data types are read from the full-process data, and timeline mapping is performed on each data type. Each time point on the timeline corresponds to several data types.

[0077] Secondly, based on the timeline and full-process data, a digital twin model is generated. The digital twin model includes physical level information, reaction kinetic level information, state estimation level information, and decision level information. Among them, the physical level information describes the measurable physical quantities in the detection system and their changes, including temperature field information, sample addition information, and equipment parameter information.

[0078] The reaction kinetics level information describes the evolution of optical signals depicting the reaction process between endotoxin and Limulus amebocyte lysate (LAL) reagent, including the original kinetic curves for each well site and derived features. The state estimation level information describes the posterior estimates of latent variables from the Bayesian state-space model, including the posterior mean, standard deviation, and confidence interval of the true endotoxin concentration over time; posterior estimates and uncertainties of temperature drift, dilution error, β-glucan interference intensity, and edge volatility factor; and convergence diagnostic metrics for particle filtering.

[0079] Decision-level information refers to the intervention decisions and their execution status that are automatically triggered during the detection process, including the time point when the online redilution command is triggered, the target command parameters; the time point when the well repositioning command is triggered, the coordinates of the moved well position; and the basis for the decision.

[0080] The information from the four levels is then structured and integrated using a timeline as an index to form a digital twin model that supports multi-perspective and multi-level access.

[0081] Furthermore, the system allows for real-time interaction with the digital twin model through a user interface. It dynamically displays the status of each moment in the detection process according to the timeline, launches independent view windows, loads the digital twin model, and drives the timeline to move from the starting point to the end point based on the user's operation commands, while continuously refreshing each visualization component.

[0082] Ultimately, if the test results determine that the endotoxin concentration exceeds the pharmacopoeia or internal control standard and the upper limit of the posterior confidence interval exceeds the limit, then the deviation point is found through a digital twin model. The deviation point is the specific location in the time-parameter space where the difference between the sample exceeding the standard and the normal sample exceeds the preset threshold.

[0083] Then, by tracing back along the detection timeline, the earliest occurrence position corresponding to the deviation point is determined, identifying the time point when the deviation exceeds the normal range. At the earliest occurrence position, the current values ​​of all parameters are listed and compared item by item with the dynamic normal range established for normal samples at the same time. Parameters with significant deviations are marked, and one or more root cause parameters are identified. The output may include, for example, parameter name, deviation time point, deviation magnitude, and recommended corrective measures.

[0084] The method provided in this application embodiment also includes: Based on the verified detection results, the node weights and statistical confidence levels in the multi-parameter knowledge graph are updated; wherein, the verification conditions include all hard constraint nodes being satisfied, Bayesian inference converging, and no unexplainable anomalies being found during virtual replay.

[0085] In this embodiment, the validation validity conditions are a set of strict criteria for screening high-quality detection results. Only detection results that meet all the conditions are considered valid and used to update the knowledge graph.

[0086] Among these conditions, "all hard constraint nodes satisfied" means that no Class I nodes were violated during the detection process; "Bayesian inference convergence" means that the particle filter algorithm reached a stable state at the end of the detection; and "no inexplicable anomalies found during virtual playback" means that no anomalies that cannot be explained by known physical, chemical, or equipment failures were found during the virtual playback of the digital twin model. Once these three conditions are met—all hard constraint nodes satisfied, Bayesian inference converged, and no inexplicable anomalies found during virtual playback—further updates can proceed.

[0087] Specifically, after the detection task is completed, based on the verified detection results, the weights of each edge pointing from a node to the detection pass rate are updated. After the update, the new node weights and confidence scores are stored in the knowledge graph database, and index reconstruction is triggered to ensure that subsequent queries can immediately use the latest weights. Subsequently, the real-time estimated sequences or final statistical values ​​of each latent variable observed during the detection process are extracted, and the data points are integrated into the historical statistics of the corresponding nodes in the knowledge graph. The mean, standard deviation, and frequency are recalculated, and the confidence interval is updated.

[0088] After the update is completed, the corresponding nodes in the knowledge graph are updated. The updated node weights and statistical confidence scores are then put into the Bayesian state-space model to set the prior distribution of latent variables for the next batch of detections, thus enabling the model to self-iterate.

[0089] In this embodiment, the satisfaction status of hard constraint nodes is monitored in real time, and the probability of violation is predicted based on the posterior estimate and its confidence interval. Intervention is automatically triggered, and when it is predicted that any hard constraint node will not be satisfied, an online re-dilution command or well repositioning command is automatically triggered to improve the reliability of decision-making and increase the success rate of single detection. Subsequently, the entire process data is continuously recorded, and virtual replay is performed based on the entire process data. When the detection shows that the standard is exceeded, the root cause parameter of the exceedance is located through virtual replay, realizing the automatic location of the root cause of the exceedance and shortening the investigation time of the exceedance. Finally, based on the verified and effective detection results, the node weights and statistical confidence in the multi-parameter knowledge graph are updated to realize the dynamic adjustment of node weights and statistical confidence, and realize the adaptive iterative update of the knowledge graph.

[0090] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment of the application, the sample to be tested is first obtained as the input parameter for the subsequent multi-parameter knowledge graph, providing a data foundation for the subsequent process.

[0091] Secondly, the input parameters are encoded into query vectors, and path retrieval and similarity matching are performed in a multi-parameter knowledge graph to obtain the most similar historical successful cases for the current sample. The detection pass rate of matching historical cases is used as the prior distribution for Bayesian optimization. The optimal solution output after iterative evaluation has a much higher expected pass rate than random selection or fixed procedures, reducing invalid detections and improving the success rate of detection solutions. Through the Bayesian optimization model, the expected uncertainty estimate of the solution detection result is given based on the posterior probability distribution, realizing efficient search in high-dimensional parameter space.

[0092] Furthermore, during the execution of the prior detection scheme, multi-source signals are collected in real time to achieve perception of the detection process and provide synchronous process observation data for the subsequent Bayesian state-space model.

[0093] Furthermore, by defining a state vector containing multiple state variables, an observation equation is established that represents the dynamic curve signal as a function of the state variables. This enables Bayesian inference to infer the temporal evolution of each latent variable, achieving accurate separation of multi-source interference. The particle filtering algorithm approximates the state space with a large number of particles, handling nonlinear and non-Gaussian state transitions and observation models, as well as the contribution of real-time interference, thus improving the accuracy of endotoxin concentration estimation. The Bayesian inference results directly provide the posterior probability distribution, extracting intervals at arbitrary confidence levels, providing reliable statistical basis for hard constraint monitoring and compliance judgment.

[0094] Finally, the system monitors the satisfaction status of hard constraint nodes in real time and predicts the violation probability based on posterior estimates and their confidence intervals, automatically triggering interventions. When it predicts that any hard constraint node will not be satisfied, it automatically triggers online reflux or well relocation commands, improving the reliability of decisions and increasing the success rate of single detections. Subsequently, it continuously records the entire process data and performs virtual replay based on the data. When a violation is detected, the root cause parameters of the violation are located through virtual replay, achieving automatic location of the root cause and shortening the investigation time. Finally, based on the verified and effective detection results, the node weights and statistical confidence in the multi-parameter knowledge graph are updated, achieving dynamic adjustment of node weights and statistical confidence, and realizing adaptive iterative updates of the knowledge graph.

[0095] Example 2, as Figure 2 As shown, based on the same inventive concept as the multi-parameter fusion endotoxin detection scheme optimization method provided in Embodiment 1, this embodiment of the invention also provides a multi-parameter fusion endotoxin detection scheme optimization system, the system comprising: The input parameter acquisition module 11 is used to load the input parameters of the sample to be tested. The input parameters include sample type, estimated endotoxin concentration range, sample matrix information and historical interference records. The test scheme generation module 12 is used to generate a priori test scheme based on the input parameters and the pre-constructed multi-parameter knowledge graph. The priori test scheme includes a recommended test method, a recommended Limulus amebocyte lysate (LAL) reagent sensitivity, a recommended dilution gradient matrix, a recommended well plate layout scheme, and an estimated value of the expected uncertainty. The multi-parameter knowledge graph contains hard constraint nodes specified in the pharmacopoeia. The multi-source signal acquisition module 13 is used to acquire multi-source signals in real time during the detection process according to the prior detection scheme. The multi-source signals include dynamic curve data of each well position, real-time temperature data of each channel, well plate position information and sample volume monitoring data. The posterior estimation module 14 is used to construct a Bayesian state-space model with the multi-source signal as input, and to estimate the temperature drift, edge volatilization effect and β-glucan false positive interference as latent variables in real time to obtain the posterior endotoxin concentration estimate and its confidence interval. The instruction triggering module 15 is used to monitor the satisfaction status of the hard constraint nodes specified in the pharmacopoeia in real time, using the estimated posterior endotoxin concentration as input, and automatically trigger an online redilution instruction or a well rearrangement instruction when it is predicted that any hard constraint node will not be satisfied.

[0096] In one embodiment, the verification scheme generation module 12 is used for: Obtain the hard constraint rules specified in the pharmacopoeia and convert them into Class I nodes. The hard constraint rules include at least the maximum effective dilution factor constraint, the sensitivity qualified interval constraint, the recovery rate range specified in the pharmacopoeia constraint, the correlation coefficient of the standard curve threshold constraint specified in the pharmacopoeia constraint, and the pH value after mixing range specified in the pharmacopoeia constraint. The soft constraint attributes in historical test records are obtained and transformed into second-type nodes. The soft constraint attributes include at least sample type, batch-to-batch variation coefficient of Limulus amebocyte lysate reagent, applicable scenario label of test method, and historical deviation data of instrument channel. Establish the association between the first type of nodes and the second type of nodes to form a multi-layered knowledge graph structure.

[0097] In one embodiment, the verification scheme generation module 12 is further configured to: The input parameters are encoded into query vectors, and the multi-parameter knowledge graph is called to perform path retrieval. Historical cases that have a similarity to the input parameters that meet a preset standard are matched as matching results. Based on the matching results, the optimal solution is searched in the preset candidate space of solutions using a Bayesian optimization model. The detection pass rate of the historical cases is used as a priori, and the prior detection solution is output after iterative evaluation.

[0098] In one embodiment, the posterior estimation module 14 is used for: Define a state vector, which includes the state variables of endotoxin true concentration, temperature deviation, dilution error, β-glucan interference intensity, and edge volatility effect. Establish an observation equation, and express the measured dynamic curve signal as a function of the state variable and the sum of the observation noise; The particle filter algorithm is used to perform Bayesian inference on the state space model. The posterior distribution is approximated through state transition and weight resampling iteration, and the contribution of each latent variable is separated in real time. Based on the Bayesian inference results, the posterior endotoxin concentration estimate and the confidence interval of the estimate are output.

[0099] The prior distribution of each latent variable in the Bayesian state-space model is set based on the historical fluctuation range and fluctuation frequency recorded by the corresponding node in the multi-parameter knowledge graph.

[0100] In one embodiment, the instruction triggering module 15 is used to: A multi-objective optimization problem is constructed, with the objective function being to minimize the total uncertainty. The decision variables include the dilution factor, the sensitivity of the horseshoe crab reagent, the sample volume, and the incubation time. The constraints include all hard constraint nodes. The second generation of the non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem and generate a Pareto optimal front solution set. Based on the Pareto optimal frontier, the optimal solution that satisfies the current detection conditions is selected. The optimal solution includes the optimized values ​​for dilution factor, Limulus amebocyte lysate (LAL) reagent sensitivity, sample loading volume, and incubation time. The optimal dilution factor in the optimal solution is compared with the current dilution factor. If the optimal dilution factor is greater than the current dilution factor, an online re-dilution command is triggered to supplement the dilution of the sample according to the optimal dilution factor. The well plate layout parameters in the optimal solution are compared with the current well position. If the correlation coefficient of the standard curve is predicted to be lower than the threshold specified in the pharmacopoeia, a well position rearrangement instruction is triggered to move the sample to the target well position corresponding to the well plate layout parameters in the Pareto optimal frontier.

[0101] In one embodiment, the instruction triggering module 15 is further configured to: During the detection process, the entire process data is continuously recorded. The entire process data includes real-time temperature sequences at each time point, dynamic curve data of each well, timestamps of each step, sample volume data, equipment parameter settings, and state vector estimates at each time point. After the test is completed, a digital twin model is constructed based on the full-process data and a virtual replay is performed. If the test result is out of standard, the root cause parameter causing the out-of-standard result is located through the virtual replay.

[0102] Specifically, a digital twin model is constructed and virtual playback is performed. If the detection result is above the standard, the root cause parameters leading to the above-standard result are located through the virtual playback, including: A timeline for the corresponding detection task is constructed based on the full-process data. Based on the timeline and the full-process data, a digital twin model is generated, which includes physical level information, reaction dynamics level information, state estimation level information and decision level information. The digital twin model is presented virtually through an interactive interface; If the test result is out of standard, the deviation point is found by comparing the digital twin model of the normal sample that was successfully tested in the same batch, and the earliest occurrence position corresponding to the deviation point is deduced from the detection time axis. Then, the parameters of the earliest occurrence position are compared with the normal range item by item to determine one or more root cause parameters.

[0103] In one embodiment, the instruction triggering module 15 is further configured to: Based on the verified detection results, the node weights and statistical confidence levels in the multi-parameter knowledge graph are updated; wherein, the verification conditions include all hard constraint nodes being satisfied, Bayesian inference converging, and no unexplainable anomalies being found during virtual replay.

[0104] Compared to existing technologies, this application first obtains the sample to be tested, which serves as the input parameter for the subsequent multi-parameter knowledge graph, thus providing a data foundation for the process.

[0105] Secondly, the input parameters are encoded into query vectors, and path retrieval and similarity matching are performed in a multi-parameter knowledge graph to obtain the most similar historical successful cases for the current sample. The detection pass rate of matching historical cases is used as the prior distribution for Bayesian optimization. The optimal solution output after iterative evaluation has a much higher expected pass rate than random selection or fixed procedures, reducing invalid detections and improving the success rate of detection solutions. Through the Bayesian optimization model, the expected uncertainty estimate of the solution detection result is given based on the posterior probability distribution, realizing efficient search in high-dimensional parameter space.

[0106] Furthermore, during the execution of the prior detection scheme, multi-source signals are collected in real time to achieve perception of the detection process and provide synchronous process observation data for the subsequent Bayesian state-space model.

[0107] Furthermore, by defining a state vector containing multiple state variables, an observation equation is established that represents the dynamic curve signal as a function of the state variables. This enables Bayesian inference to infer the temporal evolution of each latent variable, achieving accurate separation of multi-source interference. The particle filtering algorithm approximates the state space with a large number of particles, handling nonlinear and non-Gaussian state transitions and observation models, as well as the contribution of real-time interference, thus improving the accuracy of endotoxin concentration estimation. The Bayesian inference results directly provide the posterior probability distribution, extracting intervals at arbitrary confidence levels, providing reliable statistical basis for hard constraint monitoring and compliance judgment.

[0108] Finally, the system monitors the satisfaction status of hard constraint nodes in real time and predicts the violation probability based on posterior estimates and their confidence intervals, automatically triggering interventions. When it predicts that any hard constraint node will not be satisfied, it automatically triggers online reflux or well relocation commands, improving the reliability of decisions and increasing the success rate of single detections. Subsequently, it continuously records the entire process data and performs virtual replay based on the data. When a violation is detected, the root cause parameters of the violation are located through virtual replay, achieving automatic location of the root cause and shortening the investigation time. Finally, based on the verified and effective detection results, the node weights and statistical confidence in the multi-parameter knowledge graph are updated, achieving dynamic adjustment of node weights and statistical confidence, and realizing adaptive iterative updates of the knowledge graph.

Claims

1. A method for optimizing a multi-parameter fusion endotoxin detection scheme, characterized in that, include: Load the input parameters of the sample to be tested, including sample type, estimated endotoxin concentration range, sample matrix information and historical interference records; Based on the input parameters and the pre-constructed multi-parameter knowledge graph, a priori detection scheme is generated. The priori detection scheme includes a recommended detection method, a recommended Limulus amebocyte lysate (LAL) reagent sensitivity, a recommended dilution gradient matrix, a recommended well plate layout scheme, and an estimated uncertainty value. The multi-parameter knowledge graph contains hard constraint nodes specified in the pharmacopoeia. During the detection process according to the prior detection scheme, multi-source signals are acquired in real time. The multi-source signals include dynamic curve data of each well position, real-time temperature data of each channel, well plate position information and sample volume monitoring data. Using the multi-source signals as input, a Bayesian state-space model is constructed. Temperature drift, edge volatilization effect and β-glucan false positive interference are used as latent variables for real-time estimation to obtain the estimated posterior endotoxin concentration and its confidence interval. Using the estimated endotoxin concentration as input, the system monitors in real time the satisfaction status of the hard constraint nodes specified in the pharmacopoeia, and automatically triggers an online redilution command or a well rearrangement command when it predicts that any hard constraint node will not be satisfied.

2. The method as described in claim 1, characterized in that, Before loading the input parameters of the sample to be tested, the process also includes the step of constructing the multi-parameter knowledge graph: Obtain the hard constraint rules specified in the pharmacopoeia and convert them into Class I nodes. The hard constraint rules include at least the maximum effective dilution factor constraint, the sensitivity qualified interval constraint, the recovery rate range specified in the pharmacopoeia constraint, the correlation coefficient of the standard curve threshold constraint specified in the pharmacopoeia constraint, and the pH value after mixing range specified in the pharmacopoeia constraint. The soft constraint attributes in historical test records are obtained and transformed into second-type nodes. The soft constraint attributes include at least sample type, batch-to-batch variation coefficient of Limulus amebocyte lysate reagent, applicable scenario label of test method, and historical deviation data of instrument channel. Establish the association between the first type of nodes and the second type of nodes to form a multi-layered knowledge graph structure.

3. The method as described in claim 1, characterized in that, Based on the input parameters and the pre-constructed multi-parameter knowledge graph, a priori detection scheme is generated, including: The input parameters are encoded into query vectors, and the multi-parameter knowledge graph is called to perform path retrieval. Historical cases that have a similarity to the input parameters that meet a preset standard are matched as matching results. Based on the matching results, the optimal solution is searched in the preset candidate space of solutions using a Bayesian optimization model. The detection pass rate of the historical cases is used as a priori, and the prior detection solution is output after iterative evaluation.

4. The method as described in claim 1, characterized in that, A Bayesian state-space model was constructed, and temperature drift, edge volatilization effect, and β-glucan false positive interference were used as latent variables for real-time estimation to obtain posterior endotoxin concentration estimates and their confidence intervals, including: Define a state vector, which includes the state variables of endotoxin true concentration, temperature deviation, dilution error, β-glucan interference intensity, and edge volatility effect. Establish an observation equation, and express the measured dynamic curve signal as a function of the state variable and the sum of the observation noise; The particle filter algorithm is used to perform Bayesian inference on the state space model. The posterior distribution is approximated through state transition and weight resampling iteration, and the contribution of each latent variable is separated in real time. Based on the Bayesian inference results, the posterior endotoxin concentration estimate and the confidence interval of the estimate are output.

5. The method as described in claim 1, characterized in that, The prior distribution of each latent variable in the Bayesian state-space model is set based on the historical fluctuation range and fluctuation frequency recorded by the corresponding node in the multi-parameter knowledge graph.

6. The method as described in claim 1, characterized in that, The system monitors in real time the satisfaction status of the hard constraint nodes specified in the pharmacopoeia, and automatically triggers an online redilution command or well rearrangement command when it predicts that any hard constraint node will not be satisfied, including: A multi-objective optimization problem is constructed, with the objective function being to minimize the total uncertainty. The decision variables include the dilution factor, the sensitivity of the horseshoe crab reagent, the sample volume, and the incubation time. The constraints include all hard constraint nodes. The second generation of the non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem and generate a Pareto optimal front solution set. Based on the Pareto optimal frontier, the optimal solution that satisfies the current detection conditions is selected. The optimal solution includes the optimized values ​​for dilution factor, Limulus amebocyte lysate (LAL) reagent sensitivity, sample loading volume, and incubation time. The optimal dilution factor in the optimal solution is compared with the current dilution factor. If the optimal dilution factor is greater than the current dilution factor, an online re-dilution command is triggered to supplement the dilution of the sample according to the optimal dilution factor. The well plate layout parameters in the optimal solution are compared with the current well position. If the correlation coefficient of the standard curve is predicted to be lower than the threshold specified in the pharmacopoeia, a well position rearrangement instruction is triggered to move the sample to the target well position corresponding to the well plate layout parameters in the Pareto optimal frontier.

7. The method as described in claim 1, characterized in that, Also includes: During the detection process, the entire process data is continuously recorded. The entire process data includes real-time temperature sequences at each time point, dynamic curve data of each well, timestamps of each step, sample volume data, equipment parameter settings, and state vector estimates at each time point. After the test is completed, a digital twin model is constructed based on the full-process data and a virtual replay is performed. If the test result is out of standard, the root cause parameter causing the out-of-standard result is located through the virtual replay.

8. The method as described in claim 7, characterized in that, A digital twin model is constructed and virtual playback is performed. If the detection result is out of standard, the root cause parameters leading to the out-of-standard result are located through the virtual playback, including: A timeline for the corresponding detection task is constructed based on the full-process data. Based on the timeline and the full-process data, a digital twin model is generated, which includes physical level information, reaction dynamics level information, state estimation level information and decision level information. The digital twin model is presented virtually through an interactive interface; If the test result is out of standard, the deviation point is found by comparing the digital twin model of the normal sample that was successfully tested in the same batch, and the earliest occurrence position corresponding to the deviation point is deduced from the detection time axis. Then, the parameters of the earliest occurrence position are compared with the normal range item by item to determine one or more root cause parameters.

9. The method as described in claim 1, characterized in that, Also includes: Based on the verified detection results, the node weights and statistical confidence levels in the multi-parameter knowledge graph are updated; wherein, the verification conditions include all hard constraint nodes being satisfied, Bayesian inference converging, and no unexplainable anomalies being found during virtual replay.

10. A multi-parameter fusion endotoxin detection scheme optimization system, characterized in that, The system is used to implement the multi-parameter fusion endotoxin detection scheme optimization method according to any one of claims 1-9, the system comprising: The input parameter acquisition module is used to load the input parameters of the sample to be tested. The input parameters include sample type, estimated endotoxin concentration range, sample matrix information and historical interference records. The test scheme generation module is used to generate a priori test scheme based on the input parameters and a pre-built multi-parameter knowledge graph. The priori test scheme includes a recommended test method, a recommended Limulus amebocyte lysate (LAL) reagent sensitivity, a recommended dilution gradient matrix, a recommended well plate layout scheme, and an estimated value of the expected uncertainty. The multi-parameter knowledge graph contains hard constraint nodes specified in the pharmacopoeia. The multi-source signal acquisition module is used to acquire multi-source signals in real time during the detection process according to the prior detection scheme. The multi-source signals include dynamic curve data of each well position, real-time temperature data of each channel, well plate position information and sample volume monitoring data. The posterior estimation module is used to construct a Bayesian state-space model with the multi-source signal as input, and to estimate the temperature drift, edge volatilization effect and β-glucan false positive interference as latent variables in real time to obtain the posterior endotoxin concentration estimate and its confidence interval. The instruction triggering module is used to monitor the satisfaction status of the hard constraint nodes specified in the pharmacopoeia in real time, using the estimated posterior endotoxin concentration as input, and automatically trigger an online redilution instruction or a well rearrangement instruction when it is predicted that any hard constraint node will not be satisfied.

Images