Intelligent management and control method and system for blood cell separator

Abstract

The invention relates to an intelligent management and control method and system for a blood cell separator, relates to the technical field of intelligent control of medical instruments and medical information, and solves the problems that the effectiveness of a separation scheme is insufficient and the risk of complications is increased due to the fact that individual differences of patients cannot be accurately adapted. Inputting the blood indexes of the patient into a preset deep neural network, selecting an adaptive scheme from a plurality of preset subdivision separation modes, generating an initial centrifugal parameter combination, and dynamically adjusting a centrifugal rotating speed and an acceleration parameter according to a preset frequency by adopting a model prediction control algorithm in combination with data collected in real time and a separation process dynamic model; and monitoring state variables and sensor data of the dynamic model in the separation process in real time, and when monitoring parameters deviate from a preset range, starting a parameter compensation mechanism corresponding to the abnormal type. The method has the following effects that individualized precise control and real-time optimization of blood cell separation are achieved, and safety and efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent control of medical devices and medical information technology, and in particular to an intelligent management and control method and system for a blood cell separation machine. BACKGROUND

[0002] A blood cell separation machine is a key device for clinical treatment, and its intelligent management and control involves medical data sovereignty and technical autonomy. Realizing autonomous and controllable management is of great significance to ensuring clinical safety and promoting medical innovation.

[0003] Existing management and control technologies mostly rely on preset parameters and manual experience, and although some machine learning algorithms are introduced, they still use single-modal data and static models. No comprehensive intelligent decision-making system has been formed in the aspects of pipeline pre-charging, parameter adjustment, and equipment maintenance.

[0004] The existing blood cell separation machine management and control technology has obvious limitations. On the one hand, the scheme decision mainly relies on preset parameters and manual experience, and although machine learning is introduced, it is mostly based on single-modal data and static models, which cannot accurately adapt to individual differences of patients, resulting in insufficient effectiveness of the separation scheme and increased risk of complications. On the other hand, equipment maintenance relies on regular maintenance and manufacturer technical support, and lacks an intelligent monitoring and predictive maintenance mechanism based on equipment operation data, which not only increases maintenance costs, but also poses a risk of equipment operation interruption, making it difficult to ensure the continuity and stability of medical services. SUMMARY

[0005] In order to realize individualized and accurate control and real-time optimization of blood cell separation, and improve safety and efficiency, the present application provides an intelligent management and control method and system for a blood cell separation machine.

[0006] In a first aspect, the present application provides an intelligent management and control method for a blood cell separation machine, which adopts the following technical solution:

[0007] An intelligent management and control method for a blood cell separation machine, comprising:

[0008] Starting the blood cell separation machine and executing a sensor self-checking program, importing a patient blood test report and physiological parameters, and matching an adaptive basic separation scheme from a plurality of preset separation schemes through a preset classification algorithm, while automatically configuring a disposable pipeline pre-charging scheme including pre-charging fluid types, pre-charging amounts, and anticoagulant ratios based on the patient's physiological parameters;

[0009] After completing the pipeline pre-charging based on the pre-charging scheme, real-time collection of pressure, flow, optical, and vibration data is performed using a multi-modal sensing unit at a preset sampling frequency, and after preprocessing, a data fusion algorithm is used to construct a separation process dynamic model containing a preset number of state variables;

[0010] Obtaining a patient blood index, inputting the patient blood index into a preset deep neural network, selecting an adaptive scheme from a plurality of preset subdivision separation modes and generating an initial centrifugal parameter combination, using a model predictive control algorithm combined with real-time collected data and a separation process dynamic model, dynamically adjusting the centrifugal speed and acceleration parameters at a preset frequency;

[0011] Real-time monitoring of state variables and sensor data of the separation process dynamic model, when the monitoring parameters deviate from the preset range, starting the parameter compensation mechanism corresponding to the abnormal type;

[0012] The separation process data is encrypted by a preset encryption algorithm and uploaded to a blockchain platform, and a maintenance plan is generated based on the abnormal detection result of the equipment operation data.

[0013] By using the above technical scheme, the method realizes the intelligent control of the blood cell separation machine through sensor self-checking, intelligent matching of separation schemes, real-time data acquisition and model fusion, dynamic adjustment of centrifugal parameters, abnormal compensation and blockchain data storage, effectively improves the separation precision and efficiency, enhances the stability and reliability of the equipment operation, and facilitates equipment maintenance and data tracing.

[0014] In a second aspect, the present application provides a blood cell separation machine intelligent control system, which adopts the following technical scheme:

[0015] A blood cell separation machine intelligent control system includes a memory, a processor, and a program stored on the memory and executable on the processor. The program can be loaded and executed by the processor to implement the blood cell separation machine intelligent control method as described in the first aspect. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of a blood cell separation machine intelligent control method according to an embodiment of the present application.

[0017] Figure 2 is a flowchart of another embodiment of the present application, which matches the adaptive basic separation scheme from a plurality of preset separation schemes by a preset classification algorithm. DETAILED DESCRIPTION

[0018] The present application will be further described in detail below with reference to the accompanying drawings.

[0019] Reference Figure 1 A blood cell separation machine intelligent control method disclosed by the present application includes:

[0020] Step S100, start the blood cell separator and perform the sensor self-checking program, import the patient's blood test report and physiological parameters, match the appropriate basic separation scheme from multiple preset separation schemes through the preset classification algorithm, and automatically configure the disposable pipeline pre-charging scheme including the pre-charging fluid type, pre-charging volume and anticoagulant ratio based on the patient's physiological parameters.

[0021] The sensor self-checking program: a detection process automatically run when the device starts, used to confirm that each sensor (such as pressure, flow sensor) is in normal state. It is automatically executed by the built-in system of the device. Patient blood test report: including hematocrit, hemoglobin concentration, platelet count and other indicators, reflecting the patient's blood condition. It is imported through the hospital information system (HIS) or manually input. Physiological parameters: such as weight, blood pressure, body temperature, etc., which affect the selection of separation strategy. Usually measured by medical staff and entered into the system. Classification algorithm: a machine learning model (such as decision tree, support vector machine), which classifies samples according to input features. It is used here to select the most suitable separation mode for the patient from multiple preset schemes. Basic separation scheme: a pre-set standard operating procedure, such as "plasma collection mode", "platelet enrichment mode", etc. Disposable pipeline pre-charging scheme: to avoid blood clotting and ensure pipeline cleanliness, pre-charging fluid (such as saline + anticoagulant) is injected before starting, and its type and proportion are automatically configured according to the patient's condition.

[0022] The approximate process is described as follows (by way of example): 1. After the device is powered on, the sensor self-checking program is automatically run to confirm that each sensor (such as pressure, flow sensor) is in normal working condition; 2. The medical staff import the patient's blood test data (such as hematocrit 42%), weight 75 kg, normal blood pressure and other physiological parameters; 3. The system calls the pre-trained classification algorithm model (such as using a decision tree model) for analysis to determine which basic separation mode is suitable for the patient; example: if the hematocrit is high and red blood cells need to be supplemented, then match the "red blood cell collection mode"; 4. At the same time, the system automatically configures the disposable pipeline pre-charging scheme based on the patient's weight and blood viscosity; example: for an adult weighing 75 kg, the pre-charging fluid combination is saline 60 ml + anticoagulant ACD-A 12 ml (ratio 1:5); 5. After the configuration is completed, the device enters the ready state, preparing to enter the next pre-charging and separation process.

[0023] Step S200, after completing the pipeline pre-charging based on the pre-charging scheme, use the multi-modal perception unit to collect pressure, flow, optical and vibration data in real time at a preset sampling frequency, and use the data fusion algorithm to construct a separation process dynamic model containing a preset number of state variables after preprocessing.

[0024] Among them, the multi-modal perception unit: integrates the data acquisition module of multiple sensors (such as pressure, flow, optical, vibration sensors), which is used to monitor various physical parameters in the separation process in real time. Provided by the built-in hardware system of the device. The preset sampling frequency: refers to the number of times of collecting sensor data per unit time, for example, 100 times per second. Set by the device control system. Data preprocessing: including denoising, normalization, missing value filling and other operations to improve the accuracy of subsequent modeling. Automatically executed by the embedded algorithm of the system. Data fusion algorithm: a method of integrating data from multiple sensors into a unified information representation, such as principal component analysis (PCA), Kalman filter, etc. Separation process dynamic model: a mathematical model that reflects the state change of blood during the separation process, usually composed of multiple key state variables, such as liquid flow velocity, centrifugal force, red blood cell concentration distribution, etc.

[0025] The approximate process is described as follows (for example): 1. After completing the pre-charging operation of the disposable pipeline (such as injecting a physiological saline + anticoagulant mixture), the system enters the operation preparation stage; 2. Start the multi-modal perception unit to continuously collect various sensor data at the preset sampling frequency (for example, 100Hz per second); Example: the collected pressure in the current pipeline is 15kPa, the liquid flow rate is 30ml / min, and the vibration intensity is normal; 3. Preprocess the raw data to remove noise interference and standardize the format; Use data fusion algorithms (such as using principal component analysis PCA) to integrate multi-source heterogeneous data into a unified state description; Example: the fused state variables include "effective flow rate", "liquid column stability index", "centrifugal cavity vibration coefficient", etc.; 4. Based on the above variables, build a separation process dynamic model for subsequent control and abnormal monitoring; Example: the dynamic model output shows that the current plasma flow is stable, but the red blood cell aggregation trend is slightly rising, which needs to be paid attention to the subsequent changes.

[0026] Step S300, obtain the patient's blood index, input the patient's blood index into the preset deep neural network, select the adaptive scheme from multiple preset subdivision separation modes and generate an initial centrifugation parameter combination, use the model predictive control algorithm combined with real-time collected data and the separation process dynamic model to dynamically adjust the centrifugation speed and acceleration parameters at a preset frequency.

[0027] Among them, the patient's blood indicators include red blood cell count, hemoglobin concentration, platelet content, etc., reflecting the composition of blood components. Obtained through laboratory test reports or bedside rapid detection equipment. Deep neural network (DNN): a multi-layer structure of artificial neural network model for feature extraction and classification decision of complex data. In this system, it is used to match the best separation mode. Subdivision separation mode: refined operation strategy based on the basic separation scheme, such as "high hemoglobin patient plasma priority collection", "low platelet patient precise enrichment mode" and so on. Initial centrifugation parameter combination: refers to the initial centrifugation speed (RPM) and acceleration (RCF) set, which affects the separation efficiency and cell integrity. Model predictive control algorithm (MPC): an advanced control method based on mathematical model to predict future state and optimize control input. Real-time collected data: sensor information such as pressure, flow, vibration from multi-modal perception unit. Separation process dynamic model: mathematical model describing the current separation state constructed by S200, serving as the input basis for MPC algorithm.

[0028] The approximate process is described as follows (for example): 1. The system imports the patient's blood indicators (such as hemoglobin concentration of 150 g / L, hematocrit of 40%, and low platelet count); 2. Input the data into the pre-trained deep neural network model (such as using a three-layer fully connected neural network) for analysis; example output: match subdivision separation mode A1 of "moderate plasma yield + red blood cell retention"; 3. According to the mode, the system generates the initial centrifugation parameter combination, such as setting the speed to 3000 RPM and the acceleration to 3000 x g; 4. Start the model predictive control algorithm (MPC), combined with the current real-time sensor data (such as flow rate, pressure) and dynamic model prediction trend; example rule: if the model predicts that the red blood cell aggregation degree will increase, automatically reduce the acceleration to 2800 RPM to slow down the settling speed; 5. Adjust the centrifugation parameters dynamically according to the preset frequency (such as every 10 seconds), to ensure optimal and safe separation effect.

[0029] Step S400, real-time monitoring of state variables and sensor data of separation process dynamic model, when the monitoring parameters deviate from the preset range, start the parameter compensation mechanism corresponding to the abnormal type.

[0030] Among them, the state variable: the key parameter extracted from the separation process dynamic model, used to describe the system running state, such as "liquid flow stability", "red blood cell aggregation index", "centrifugal cavity pressure fluctuation value" and so on. Sensor data: real-time acquisition data from multi-modal perception unit, including pressure, flow, optical signal, vibration, etc., provided by the built-in sensor of the device. Preset range: the normal operation interval set for each state variable, which is judged as abnormal if it exceeds. The range is pre-set based on clinical experience and equipment safety standards. Parameter compensation mechanism: a set of adjustment strategies automatically executed by the system when a parameter deviates from the normal range, to restore stable operation.

[0031] The approximate process is described as follows (for example): 1. The system continuously monitors the state variables (such as the red blood cell sedimentation rate) and sensor data (such as the pipeline pressure); 2. If a certain indicator deviates from the preset range (for example, the pressure suddenly rises to 25kPa, exceeding the set upper limit of 20kPa), trigger abnormal judgment; 3. According to the type of abnormality, match the corresponding parameter compensation mechanism; example rule: if "pressure is too high", start the "reduce pump speed + increase anticoagulant proportion" compensation scheme; 4. The compensation mechanism automatically adjusts the device parameters through the control system, such as reducing the pump speed from 30ml / min to 25ml / min, and adjusting the anticoagulant proportion from 1:5 to 1:4; 5. Continue to monitor the state variables after adjustment, if it returns to normal, maintain the current parameters; if it is still abnormal, further upgrade the compensation strategy or prompt manual intervention.

[0032] Step S500, the separation process data is encrypted by a preset encryption algorithm and uploaded to the blockchain platform, and a maintenance plan is generated based on the abnormal detection result of the device running data.

[0033] Among them, the separation process data: including the key information of pressure, flow, centrifugal parameter, blood index change collected in the whole blood cell separation process. Generated by real-time recording during device operation. Preset encryption algorithm: encryption standard for protecting data security, such as AES (Advanced Encryption Standard) or SM4 (National Encryption Algorithm). Provided by the built-in security module of the system. Blockchain platform: a distributed ledger technology platform, with characteristics of non-tamperable, traceable, etc., suitable for medical data storage and sharing. Connected with the device through API interface. Device running data: records the working state, fault log, sensor performance, etc. of the device in this and historical operation. Automatically recorded by the device control system. Abnormal detection result: potential failure or performance decline identified by analyzing the device running data, such as sensor drift, pump response delay, etc. Judged by rule engine or lightweight machine learning model (such as Isolation Forest). Maintenance plan: suggested maintenance measures generated according to the health status of the device, such as "clean the centrifugal cavity", "replace the pressure sensor", etc. Automatically generated by the system according to the type of abnormality.

[0034] The general process is illustrated as follows: 1. After the separation process is completed, the system will package the separation process data collected in this operation (such as centrifugal speed curve, red blood cell concentration change, pipeline pressure fluctuation); 2. Use the built-in encryption algorithm (such as AES-256) in the system to encrypt the data to ensure privacy and integrity; 3. Upload the encrypted data to the blockchain platform through the secure communication protocol for subsequent audit, traceability or remote monitoring; At the same time, the system calls the abnormal detection model (such as based on the isolation forest algorithm) to analyze the equipment operation data; Example output: found that the "pressure sensor response delay increases" in this operation, and judged as a potential aging problem; 4. According to the detection result, the system automatically generates a maintenance plan; Example content: "suggest replacing the pressure sensor within 7 days", and push to the hospital equipment management system; 5. Maintenance personnel can perform preventive maintenance according to the plan to improve equipment reliability and operation safety.

[0035] Referring to Figure 2 , the suitable basic separation scheme is matched from a plurality of preset separation schemes by a preset classification algorithm, including:

[0036] In step S110, the numerical features in the patient blood test report are mapped to a semantic vector space through contrastive learning, and at the same time, the physiological parameters are converted into high-dimensional embedding vectors by using a medical knowledge graph, and a patient feature tensor containing clinical semantics is formed by fusing through an attention mechanism.

[0037] Among them, the numerical features in the patient blood test report: such as red blood cell count, hemoglobin concentration, platelet number, etc. Quantitative indicators. Imported through the laboratory information system (LIS) or manual input method. Contrastive learning (Contrastive Learning): an unsupervised or weakly supervised learning method, which is used to map data into a vector space, so that similar samples are closer and dissimilar samples are farther apart. It is used here to embed numerical features into a semantic space. Medical knowledge graph: a knowledge network that expresses medical entities and their relationships in a structured way, such as "high blood pressure → affects → kidney", "anticoagulant → acts on → coagulation factor". It can be constructed by authoritative medical databases (such as SNOMED CT, UMLS). High-dimensional embedding vector: map the original data (such as text, symbol, parameter) to a vector representation in a high-dimensional space for subsequent model processing. Attention mechanism (Attention Mechanism): a mechanism in deep learning that automatically focuses on key parts in multi-source information to improve fusion effect. Patient feature tensor: the multi-modal feature representation after fusion, usually a three-dimensional or higher structure (such as [batch, features, time]), which is used as the input of the subsequent classifier.

[0038] The general process is illustrated as follows: 1. The system reads the patient's blood test report and extracts numerical features (such as hematocrit 42% and hemoglobin 150 g / L); 2. Use the pre-trained contrast learning model (such as SimCLR structure) to map these numerical features to a unified semantic vector space, so that they can be processed with other non-numerical information; Example: "hemoglobin 150 g / L" is mapped to vector [0.87, -0.32, 0.15]; At the same time, use the medical knowledge graph to encode the patient's physiological parameters (such as "weight 75 kg" and "normal blood pressure") into high-dimensional embedding vectors; Example: "normal blood pressure" corresponds to graph node ID "BP_001", and the embedding is vector [0.65, 0.12, -0.43]; 4. Use attention mechanism (such as QKV structure in Transformer) to weight and integrate the two types of vectors to generate a unified patient feature tensor; Example output tensor dimension is [1x64], where each element represents a different dimension of clinical semantic feature; 5. The fused tensor will be used as the input of the next special classifier to match the most suitable separation scheme.

[0039] Step S120, based on the model-independent meta-learning framework, the pre-trained medical field meta-model is used to perform gradient update on the patient feature tensor for a preset number of times to generate a special classifier for the current patient.

[0040] Among them, the model-independent meta-learning framework (MAML, Model-Agnostic Meta-Learning): a general meta-learning method, which is trained on multiple tasks to enable the model to quickly adapt to new tasks. Suitable for various neural network structures. Medical field meta-model: a basic model trained on a large amount of multi-center clinical data, with good generalization ability. It can be obtained through local deployment or cloud loading. Patient feature tensor: a high-dimensional tensor that combines numerical features and semantic information output by S110, used as the input of this step. Derived from the internal transmission of the device control system. Gradient update: in deep learning, the process of adjusting model parameters by calculating the gradient of the loss function. In this case, it is used to fine-tune the meta-model. Special classifier: a customized classification model for the current patient, which can more accurately match the applicable basic separation scheme.

[0041] The general process is described as follows: 1. The system loads a pre-trained meta-model in the medical field (such as a classification model based on Transformer), which has completed meta-learning training on multiple blood cell separation cases. 2. Input the patient's feature tensor (for example, a vector with dimensions [1x64]). In the MAML framework, the meta-model is updated for a limited number of times (such as 3 times) using only the feature data of the current patient for fine-tuning; Example: using the cross-entropy loss function, the learning rate is set to 0.01, and the updated model is more sensitive to the current patient's features. 3. After fine-tuning, a dedicated classifier for the patient is generated; Example output: the classifier judges that the "red blood cell collection mode" matching probability is 85%, higher than other options; The dedicated classifier will be used for subsequent steps of basic separation scheme recommendation and trust assessment.

[0042] Step S130, build a fuzzy reasoning engine through a preset number of clinical rules, combine the output results of the dedicated classifier, time series analysis model, and image recognition model using Dempster-Shafer evidence theory to calculate the trust score of each basic separation scheme.

[0043] Among them, the clinical rules: decision logic based on medical guidelines or expert experience, such as "avoid high-intensity centrifugation when platelet count is less than 50x10 9 / L". Derived from authoritative medical literature or hospital knowledge base. Fuzzy reasoning engine: a reasoning system for handling uncertain information, which can convert imprecise input into qualitative judgments. Provided by the system's built-in logic module. Dempster-Shafer evidence theory (D-S theory): a mathematical method for handling uncertain reasoning, which can be used to fuse evidence from multiple sources and calculate the overall trust. Dedicated classifier: generated by S120, optimized classification model output results for the current patient. Internal transmission from the previous step. Time series analysis model: a model for processing time series data, such as LSTM, GRU, etc., which can identify the trend of blood parameters over time. Pre-trained and deployed in the device. Image recognition model: used to identify blood images (such as red blood cell aggregation state) collected by optical sensors, such as CNN model. Also a pre-trained model. Basic separation scheme: multiple pre-set blood cell separation operation modes, such as "plasma priority collection", "red blood cell enrichment", "whole blood component separation", etc. Provided by the device standard process library.

[0044] The approximate process is described (for example) as follows: 1. The system collects the output results of the three models: the dedicated classifier predicts that the "plasma priority collection" confidence is 0.8; the time series analysis model predicts that the hematocrit is in a downward trend, supporting the scheme; the image recognition model detects that the current blood image has uniform red blood cell distribution without obvious aggregation; the above results are taken as input and are semantically converted through the fuzzy reasoning engine; an example rule is: "if the red blood cell distribution is uniform and the hematocrit is stable, then the trust weight of the 'plasma priority' scheme is enhanced"; 2. The Dempster-Shafer evidence theory is used to fuse and calculate the three types of evidence to obtain the comprehensive confidence of each separation scheme; an example output is: the final confidence of the "plasma priority collection" scheme is 0.87, and that of the "red blood cell enrichment" is 0.65; the system preliminarily filters out the most suitable basic separation scheme according to the confidence ranking and passes it to the subsequent risk assessment link.

[0045] In step S140, the classification model parameters of multiple medical institutions are aggregated through a federated learning framework to periodically update the global classifier.

[0046] The federated learning (FL) framework is a distributed machine learning method that allows multiple institutions to collaboratively train models without sharing raw data, protecting data privacy. It is supported by the built-in communication and calculation modules of the system. The classification model parameters refer to the internal parameters of the dedicated classifier used for basic separation scheme matching (such as weights and biases in neural networks), which are generated by the local device when performing steps S120 and the like. Medical institutions include hospitals, blood stations, research centers, and other units that use the blood cell separation machine. They are authorized to access the federated learning network. The global classifier is the main model deployed on the central server, which periodically aggregates local model parameter updates from each participating institution to improve overall classification performance. It is provided and managed by the system maintainer.

[0047] The approximate process is described (for example) as follows: 1. Multiple medical institutions locally run their own blood cell separation systems, and a dedicated classifier parameter (such as model weight fine-tuned for the characteristics of a certain patient) is generated each time the system is used; under the premise of ensuring patient privacy, each institution uploads the encrypted model parameter update to a unified federated learning platform; for example, a hospital uploads the update parameter ΔW1, and another center uploads ΔW2; 2. The federated learning platform uses a weighted average algorithm (such as FedAvg) to aggregate all uploaded parameters; for example, the new global parameter W_global_new = 0.5 x ΔW1 + 0.5 x ΔW2; 3. The updated global classifier is pushed back to each participating institution for initialization of the next round of dedicated classifier; the entire process is periodically executed (such as once a week) to continuously optimize the classification model and adapt to more clinical scenarios.

[0048] Step S150, evaluate the potential risks of each candidate scheme based on the causal inference model, quantify the benefit-risk ratio of the scheme through counterfactual reasoning, and select the basic separation scheme with the highest trust degree and risk indicators meeting the standards, wherein the potential risks include hemolysis probability and coagulation risk.

[0049] Wherein, the causal inference model (Causal Inference Model): a statistical modeling method for analyzing the causal relationship between variables, which can identify the impact of intervention factors on results. In this system, it is used to evaluate the impact of different separation schemes on patient risk. It is pre-trained and deployed by the system. Counterfactual reasoning (Counterfactual Reasoning): In causal inference, the potential consequences of different decisions are evaluated by simulating "what would happen if other operations were taken". Basic separation scheme: multiple preset blood cell separation operation modes, such as "plasma priority collection" and "red blood cell enrichment". Derived from the standard process library of the device. Potential risk: refers to adverse events that may occur during blood cell separation, such as: hemolysis probability: the possibility of red blood cell rupture releasing hemoglobin; coagulation risk: the probability of local or overall coagulation of blood in the pipeline. Benefit-risk ratio: an evaluation index that considers the treatment effect and safety of the scheme, used to assist in selecting the optimal scheme.

[0050] The approximate process is described as follows (by example): the system inputs multiple candidate separation schemes and their trust degrees output by S130 into the causal inference model; the model conducts counterfactual analysis combined with the patient's blood characteristics (such as hematocrit 42% and anticoagulant usage); example reasoning: "if high-speed centrifugation is used, the hemolysis probability will increase by 5%"; for each scheme, its expected efficacy (such as plasma collection efficiency) and potential risk (such as hemolysis and coagulation probability) are evaluated; based on the above results, the benefit-risk ratio of each scheme is calculated; example output: scheme A: benefit-risk ratio = 0.82 (high collection efficiency + moderate risk); scheme B: benefit-risk ratio = 0.75 (moderate collection efficiency + low risk); finally, the basic separation scheme with the highest trust degree and risk indicators meeting the standards is selected; example decision: choose scheme A, as it has higher treatment benefits under the premise of safety.

[0051] Evaluating the potential risks of each candidate scheme based on the causal inference model, quantifying the benefit-risk ratio of the scheme through counterfactual reasoning, and selecting the basic separation scheme with the highest trust degree and risk indicators meeting the standards include:

[0052] Step S151, construct a structural causal model containing patient characteristics, blood indicators, and device parameters, divide intervention variables, outcome variables, and covariates, and identify key influencing factors through backdoor criteria, wherein the intervention variable is the separation scheme, the outcome variable includes hemolysis and coagulation risk, and the covariate includes clinical confounding factors.

[0053] Causal Inference Model: a statistical and graphical model-based method to identify causal relationships between variables, rather than just correlations. Used in this system to evaluate the impact of different separation protocols on patient risk. Structural Causal Model (SCM): a mathematical model consisting of variables and their causal relationships, often represented as a directed acyclic graph (DAG). Pre-trained and deployed by the system. Treatment Variable: a factor that is artificially controlled or changed, in this case "the adopted blood cell separation protocol" (e.g. plasma collection mode, red cell enrichment mode, etc.). Outcome Variable: the output variable that is affected by the intervention, in this step including: hemolysis probability: the risk of red blood cell rupture releasing hemoglobin; coagulation risk: the possibility of local or overall clotting of blood in the pipeline. Covariates: other factors that can affect the outcome variable and need to be controlled in the analysis. In this case, including patient's clinical confounding factors (such as age, anticoagulant usage, underlying diseases, etc.). Backdoor Criterion: a rule in causal graphs used to identify which covariates should be adjusted to eliminate confounding effects, ensuring unbiased causal estimates.

[0054] The general process is described as follows (example): 1. The system constructs a structural causal model (SCM) based on the individual information of the patient (such as age 45, hematocrit 42%, weight 70 kg) and equipment operating parameters (such as centrifugal speed, pump speed); 2. In this model, three types of variables are clearly defined: intervention variable: "separation protocol A / B / C"; outcome variable: "whether hemolysis occurs" or "whether pipeline coagulation occurs"; covariates: including "anticoagulant dose", "blood viscosity", "underlying diseases (such as diabetes)" and other confounding factors; use causal graph and backdoor criterion for path analysis to identify key covariates that need to be controlled; example: if "anticoagulant dose" affects both "separation protocol selection" and "coagulation risk", it must be included as a covariate for adjustment; the constructed causal model will be used for subsequent causal effect calculation and counterfactual analysis.

[0055] Step S152, simulate intervention using do-operator, calculate the causal effect value of each protocol on risk through Bayesian structural equation model.

[0056] do-operator: In causal inference, a do-operator is used to simulate an intervention on a variable. Unlike ordinary probability conditioning, do-operators can remove natural dependencies between variables, allowing for more accurate estimates of intervention effects. Bayesian Structural Equation Model (BSEM): A method that combines Bayesian inference with structural equation modeling, suitable for handling data with uncertainty and complex causal relationships. Pre-trained and deployed by the system. Causal Effect Value: Represents the impact of an intervention (such as using a specific separation scheme) on an outcome variable (such as hemolysis or coagulation risk), usually expressed as a probability difference or relative risk ratio. Separation scheme: Refers to a variety of pre-set blood cell separation operation modes, such as "plasma priority collection", "red blood cell enrichment", etc. Derived from the device standard process library.

[0057] Risk variables: Include "hemolysis incidence", "pipeline coagulation probability" and other clinical safety indicators.

[0058] The general process is described as follows (example): 1. Based on the constructed structural causal model, the system uses do-operator to simulate different separation scheme interventions; Example: executing do (scheme = A) means forcing the application of the "plasma priority collection" scheme; 2. Input the data after intervention into the Bayesian Structural Equation Model (BSEM), which is based on historical case data and current patient characteristics for inference; 3. The model outputs the causal effect values corresponding to each separation scheme; Example results: Scheme A results in a hemolysis risk of 8.2%; Scheme B results in a hemolysis risk of 5.1%; Also assess the impact on other risk variables; Example: Scheme A causes coagulation risk to rise to 6.7%, higher than Scheme B's 3.5%; All causal effect values will serve as important basis for subsequent counterfactual analysis and comprehensive scoring.

[0059] Step S153, estimate potential traits based on patient data, calculate risk differences for different schemes through counterfactual queries.

[0060] Counterfactual Reasoning: A causal reasoning method to answer the question of "what would the outcome be if a different decision was taken at the time". Used here to evaluate the potential risk differences for the same patient under different separation protocols. Potential Traits: Refers to the patient's intrinsic physiological and pathological characteristics, such as blood viscosity, red blood cell fragility, anticoagulation ability, etc., which affect their response to different separation protocols. Estimated through historical case data and current test reports. Structural Causal Model (SCM): A variable causal relationship model constructed in S151 to support counterfactual queries. Called internally by the system. Bayesian Inference: A method of updating probability distributions based on prior knowledge and observed data, commonly used in uncertainty modeling. Used here to estimate the patient's individualized parameters. Risk Difference: Represents the change in hemolysis or coagulation risk when different separation protocols are used on the same patient, and is one of the core outputs of counterfactual reasoning.

[0061] The approximate process is described as follows (with examples): 1. The system estimates the patient's potential trait distribution based on their current blood indicators (e.g. hematocrit 40%, hemoglobin concentration 145g / L) and clinical characteristics (e.g. no history of diabetes); 2. Input this distribution into the constructed structural causal model (SCM) and combine it with the causal effect values calculated in S152; 3. Perform counterfactual queries to simulate the risk changes under different separation protocols; Example: Currently using protocol A, the predicted hemolysis risk is 8%; Counterfactual reasoning: "If protocol B is used", the hemolysis risk will drop to 5%; 4. Output the risk differences between each protocol as the basis for comprehensive evaluation; Example result: Protocol B is 3 percentage points lower than protocol A, and the coagulation risk is reduced by 2 percentage points.

[0062] Step S154, using Dempster-Shafer theory to fuse causal effects, fuzzy reasoning confidence, historical case data according to pre-set weights to construct a three-dimensional evaluation matrix to calculate the comprehensive score.

[0063] Dempster-Shafer evidence theory (D-S theory): a mathematical method for handling uncertain reasoning, which can be used to fuse multiple sources of inaccurate or uncertain information and calculate their trust degree on a certain conclusion. Provided by the system's built-in algorithm module. Causal effect: calculated in S152 through the Bayesian structural equation model, representing the degree of influence of a certain separation scheme on the risk of hemolysis, coagulation, etc. Derived from the output of the previous step. Fuzzy reasoning trust degree: obtained after fusing the multi-model output in S130 through the fuzzy reasoning engine and D-S theory. From the system's internal transmission. Historical case data: clinical outcome data of similar patients using different separation schemes in the past, used to assist current decision-making. Derived from the hospital database or cloud knowledge base. Three-dimensional evaluation matrix: a scoring matrix composed of three types of indicators, causal effect, trust degree, and historical efficacy, used to support the final scheme optimization.

[0064] The approximate process is described as follows (by example): 1. The system collects information in three dimensions: causal effect value: for example, scheme A leads to a hemolysis risk of 8%, and scheme B leads to a hemolysis risk of 5%; fuzzy reasoning trust degree: for example, scheme A has a trust degree of 0.87, and scheme B has a trust degree of 0.79; historical case data: for example, among the previous 100 similar patients, 78 cases were successful using scheme A, and 65 cases were successful using scheme B; 2. Normalize and unify the dimensions of the three types of information; 3. Use Dempster-Shafer evidence theory to perform weighted fusion according to the preset weights (such as 40% for causal effect, 30% for trust degree, and 30% for historical data); example rule: if a certain scheme has low risk, high trust degree, and high historical success rate in terms of causal effect, the comprehensive score will be higher; 4. Construct a three-dimensional evaluation matrix, with each dimension representing an evaluation dimension, and each candidate scheme corresponding to a comprehensive score; example output: scheme A scores 0.83, and scheme B scores 0.76.

[0065] Step S155, according to the preset safety threshold and benefit threshold, filter the schemes that meet the standard and have the optimal risk-benefit ratio.

[0066] Safety Threshold: upper limit of risk set for patient safety, such as hemolysis rate should not exceed 10%, coagulation risk should be lower than 8%. Derived from clinical guidelines or device safety specifications. Efficacy Threshold: minimum standard to measure treatment effect, for example, plasma collection efficiency should not be less than 70%. Derived from medical expert consensus or historical data statistics. Risk-Benefit Ratio: an index that comprehensively evaluates the risk accompanied by the curative effect of a certain separation scheme, usually expressed in the form of "efficiency score / risk score". Basic separation scheme: a variety of preset blood cell separation operation modes, such as "plasma priority collection", "red blood cell enrichment", etc. From the device process library.

[0067] The approximate process is described as follows (for example): 1. The system reads the comprehensive score of each candidate scheme output by S154 and the corresponding risk-benefit analysis result; according to the preset safety threshold and efficacy threshold, screening is carried out; example rules: if the hemolysis risk of a certain scheme is >10%, it is directly excluded; if the plasma collection efficiency is <70%, it is not recommended to use; calculate the risk-benefit ratio of the scheme that passes the threshold screening;

[0068] Example: Scheme A: efficiency score is 0.9, risk score is 0.08→risk-benefit ratio=0.089; Scheme B: efficiency score is 0.8, risk score is 0.05→risk-benefit ratio=0.062;

[0069] Finally, the scheme with the lowest risk and the optimal risk-benefit ratio is selected; example decision: although the efficiency of scheme A is slightly higher, the system selects scheme B because it is more optimal in safety and cost performance; output the selected scheme and push it to the control system to enter the execution phase.

[0070] According to the preset safety threshold and efficacy threshold, the scheme that meets the standard and has the optimal risk-benefit ratio includes:

[0071] Step S155.1, set the safety threshold: hemolysis risk probability <5%, coagulation risk probability <8%, and set the efficacy threshold: separation efficiency >90%, recovery rate >85%.

[0072] Safety Threshold: upper limit of risk set for patient safety. In this system, it includes:

[0073] Hemolysis risk probability <5%: indicates that the probability of red blood cell rupture releasing hemoglobin should be controlled below 5%;

[0074] Coagulation risk probability <8%: indicates that the possibility of local or overall coagulation of blood in the pipeline should be lower than 8%;

[0075] Acquisition method: Based on clinical medical guidelines (such as "Transfusion Technology Standards"), equipment safety test data and expert consensus setting.

[0076] Efficacy Threshold: The minimum standard for measuring separation effectiveness. In this system, it includes:

[0077] Separation efficiency > 90%: refers to the proportion of successful separation of target blood cell components from whole blood;

[0078] Recovery rate > 85%: refers to the proportion of the final target component obtained in the total amount of the component in the original blood;

[0079] Acquisition method: derived from clinical efficacy evaluation criteria, historical operation data analysis and equipment performance verification experiment.

[0080] Pre-set rules: the above threshold values are embedded in the control system logic as hard constraints, used to screen qualified candidate separation schemes.

[0081] Step S155.2, traverse all candidate schemes, and screen the feasible solution set that meets the safety threshold and the benefit threshold at the same time.

[0082] Candidate scheme: refers to multiple blood cell separation operation modes generated in the previous step through causal reasoning and comprehensive evaluation, such as "plasma collection first", "red blood cell enrichment", "whole blood component synchronous separation", etc. Derived from the built-in process library of the system. Safety threshold: including hemolysis risk <5%, coagulation risk <8%, used to exclude high-risk schemes. Preset by the system and can be updated regularly. Efficacy threshold: including separation efficiency > 90%, recovery rate > 85%, used to ensure that the treatment effect meets the standard. Set by the device performance standard and clinical guidelines. Feasible solution set (Feasible Solution Set): a set of candidate schemes that meet all safety and benefit constraints, serving as the basis for the next step of multi-objective optimization.

[0083] Step S155.3, multi-objective optimization analysis is performed on the schemes in the feasible solution set, and the Pareto optimal solution set is identified.

[0084] Feasible Solution Set: The set of candidate separation schemes that satisfy both safety and benefit thresholds, filtered by S155.2. Originated from internal process transmission. Multi-objective Optimization: An optimization method that seeks the optimal compromise between multiple conflicting objectives, such as both low risk and high efficiency. In this system, the objectives are "minimize risk" and "maximize benefit". Pareto Optimal Set: A set of solutions in multi-objective optimization that cannot be further improved in one objective without compromising the other. Also known as the non-inferior solution set. Calculated by the system through optimization algorithms. NSGA-II algorithm (Non-dominated Sorting Genetic Algorithm II): A widely used multi-objective genetic optimization algorithm suitable for the scheme optimization task in this system. Provided by the system's built-in optimization module.

[0085] Step S155.4, calculate the risk-benefit ratio for each scheme in the Pareto Optimal Set, and the specific risk-benefit ratio calculation formula is as follows:

[0086] Where w k is the benefit index weight, and v l is the risk index weight.

[0087] Pareto Optimal Set: A set of non-inferior solutions in multi-objective optimization that cannot be further improved in one objective without compromising the other. Originated from the output of S155.3. Risk-Benefit Ratio: A comprehensive indicator used to measure the risk a scheme undertakes while bringing therapeutic benefits, usually expressed in the form of "normalized risk score / normalized benefit score". Benefit Index Weight: Indicates the importance of separation effect in the final decision, with a value range of [0, 1], and the default value can be set to 0.5. Derived from clinical rules or system presets. Risk Index Weight: Indicates the importance of safety requirements, usually satisfying α+β=1. Also derived from system configuration. Normalization: Scaling the original index value to a unified dimension (such as between 0 and 1) for easy comparison and weighted fusion.

[0088] Step S155.5, dynamically adjust the threshold weight according to the clinical scene, and select the scheme with the maximum RRR i as the final output.

[0089] Clinical Scenario: refers to the specific treatment context that the current patient is in, such as "plasma exchange", "platelet collection" or "high-risk patients with hemolytic tendency". Derived from patient blood test reports and physiological parameters. Threshold Weight: a parameter used to adjust the proportion of risk and benefit in the comprehensive score, for example, to increase the safety weight in high-risk patients. Preset by the system according to clinical rules or dynamically configured by the doctor side. Risk-benefit ratio: calculated in S155.4, indicating the risk cost of a candidate scheme in bringing therapeutic effect. The lower the value, the higher the cost performance of the scheme. Final output scheme: the most suitable blood cell separation operation mode for the patient selected after multi-dimensional evaluation, such as "standard plasma collection mode", "low-speed red blood cell enrichment mode", etc. Derived from system decision output.

[0090] At the same time, based on the patient's physiological parameters, the disposable tubing pre-charging scheme including pre-charging fluid type, pre-charging volume and anticoagulant ratio is automatically configured, which includes:

[0091] Step S110, build a causal knowledge graph containing patient characteristics, blood indicators and pre-charging parameters, and generate a patient feature tensor using a graph neural network.

[0092] Among them, patient characteristics: including age, weight, blood pressure, blood type and other individual physiological information. Obtained through electronic medical record system (EMR) or medical staff input. Blood indicators: such as hematocrit, hemoglobin concentration, platelet count, etc., reflecting the current blood state of the patient. Derived from laboratory test reports or rapid bedside detection equipment. Pre-charging parameters: refer to key parameters involved in the pre-charging process of disposable tubing, including: pre-charging fluid type (such as normal saline, lactated Ringer's solution); pre-charging volume (such as 50ml); anticoagulant ratio (such as ACD-A to blood ratio of 1:5); derived from built-in process library and clinical operation specification. Causal knowledge graph (Causal Knowledge Graph): a structured knowledge representation method that integrates causal relationships in the medical field, used to model the causal effects between patient characteristics, blood state and pre-charging scheme. Built from authoritative medical literature, expert experience and historical case data. Graph neural network (Graph Neural Network, GNN): a deep learning model for processing graph-structured data that can extract high-level semantic features from causal knowledge graphs. In this system, it is used to map patient information into a unified feature tensor. Patient feature tensor: a multi-dimensional vector output by GNN, which integrates the potential causal relationships between patient characteristics, blood indicators and pre-charging schemes, serving as input for subsequent configuration models.

[0093] Step S120, based on the model-independent meta-learning framework, use the pre-trained meta-model to perform gradient updates on the patient feature tensor for a preset number of times to generate a dedicated configuration model.

[0094] Among them, the model-independent meta-learning framework (MAML, Model-Agnostic Meta-Learning): a general meta-learning method, by training the base model on multiple tasks, so that it has the ability to quickly adapt to new tasks. Applicable to various neural network structures. In this system, it is used to build a generalizable pre-priming scheme configuration model. Pre-trained meta-model: a pre-trained base model on a large amount of multi-institutional clinical data and historical configuration cases, with the ability to make reasonable pre-priming recommendations for different patient characteristics. Loaded by the system when deployed. Patient feature tensor: a high-dimensional vector representation that combines patient physiological parameters, blood indicators, and causal relationships output by S110, used as input in this step. Derived from internal system processes. Gradient update: a process in deep learning that adjusts model parameters through backpropagation. In this case, it is used to personalize and fine-tune the meta-model. Special configuration model: a pre-priming scheme generation model optimized for the current patient, outputting parameters such as pre-priming fluid type, pre-priming volume, and anticoagulant ratio.

[0095] The general process is described as follows (with examples): the system loads a trained meta-learning model (such as a meta-model based on a fully connected network), which has been pre-trained in multiple blood cell separation scenarios; input the patient's feature tensor (for example, a vector with a dimension of

[64] ); under the MAML framework, perform a limited number of gradient updates (such as 2 times) on the model, using only the current patient's feature data for fine-tuning; example: using the mean square error loss function, the learning rate is set to 0.005, and the updated model is more suitable for the patient's weight, hematocrit, and other characteristics; after fine-tuning, a special configuration model for the patient is generated; example output: recommend "saline + ACD-A anticoagulant" combination, pre-priming volume is 60ml, anticoagulant ratio is 1:5; the model output will be used as input for the next step of fuzzy reasoning engine, to continue participating in the dynamic optimization of pre-priming parameters.

[0096] Step S130, build a fuzzy reasoning engine through a preset number of clinical rules, combine the Dempster-Shafer evidence theory to fuse the output results of the special configuration model, the hemodynamic prediction model, and the equipment compatibility data, and dynamically adjust the pre-priming parameters.

[0097] Among them, Clinical Rules: Decision logic based on medical guidelines or expert experience, such as "Patients with body weight > 70 kg should increase the pre-charge volume" and "Patients with high hematocrit need to increase the proportion of anticoagulant". Derived from authoritative medical literature or hospital knowledge base. Fuzzy Inference Engine: A reasoning system for handling uncertainty and fuzzy input, which can convert imprecise clinical judgments into operational numerical outputs. Provided by the system's built-in logic module. Dempster-Shafer Evidence Theory (D-S Theory): A mathematical method for handling uncertain reasoning, which can be used to fuse evidence from multiple sources and calculate their combined trustworthiness. Special configuration model: generated by S120, customized optimized pre-charge parameter recommendation model output results for the current patient. Internal transmission from the previous step. Hemodynamic prediction model: a model based on fluid mechanics principles and machine learning, used to predict blood flow status (such as pressure distribution, shear force) under different pre-charge schemes. Pre-trained and deployed in the device. Device compatibility data: refers to the technical limitation conditions of the blood cell separator currently used for pre-charge liquid type, volume, flow rate, etc. Derived from the technical specifications provided by the device manufacturer.

[0098] The approximate process is described as follows (for example):

[0099] 1. The system collects three types of input information: special configuration model output: "Suggest pre-charge volume 60ml + ACD-A anticoagulant + proportion 1:5"; hemodynamic prediction model output: "Under this scheme, the maximum shear stress in the pipeline is 12Pa, which is within the safe range"; device compatibility restrictions: "The maximum allowed pre-charge volume is 80ml, and ACD-A anticoagulant is supported"; use the above information as input, and perform semantic conversion through the fuzzy inference engine; example rule: "If the hematocrit is high and the shear stress is close to the upper limit, then moderately reduce the pump speed and increase the anticoagulant proportion".

[0100] 2. Use Dempster-Shafer Evidence Theory to fuse and calculate the three types of evidence, and obtain the trustworthiness distribution of each pre-charge parameter; example output: the comprehensive trustworthiness of "anticoagulant proportion 1:5" is 0.82, and "1:4" is 0.75;

[0101] 3. According to the fusion results, dynamically adjust the initial pre-charge parameters; example adjusted output: pre-charge liquid type: physiological saline + ACD-A mixed solution; pre-charge volume: 65ml; anticoagulant ratio: 1:4; the adjusted pre-charge scheme will enter the subsequent risk assessment process (S150) to ensure safety and effectiveness.

[0102] Step S140, aggregate model parameters of multiple medical institutions through federated learning framework, periodically update global model, and compress to edge device using knowledge distillation technology.

[0103] Federated Learning Framework: A distributed machine learning method that allows multiple medical institutions to collaboratively train a global model without sharing raw patient data. Supported by the system's built-in communication module. Classification / configuration model parameters: Refers to the internal parameters of the dedicated configuration model used for pre-charge scheme recommendation (such as weights and biases in neural networks), generated by the local device when performing steps such as S120. Medical institutions: Units such as hospitals, blood banks, and research centers that use the blood cell separation system, authorized to access the federated learning network. Global Model: The main model deployed on the central server, regularly aggregates local model updates from participating institutions to improve overall generalization. Provided and managed by the system maintainer. Knowledge Distillation: A model compression technique that allows a small model (student model) to mimic the behavior of a large model (teacher model) while maintaining performance and reducing model size for deployment on edge devices. Edge Device: Refers to the blood cell separation machine control system actually running in the hospital or blood collection site, with limited resources, requiring a lightweight model.

[0104] The general process is described as follows: Multiple medical institutions run their own blood cell separation systems locally, generating a dedicated configuration model parameter (such as model weights fine-tuned for a patient's characteristics) each time they are used; under the premise of ensuring patient privacy, each institution uploads encrypted model parameter updates to a unified federated learning platform; Example: a hospital uploads update parameters ΔW1, another center uploads ΔW2; the federated learning platform uses a weighted average algorithm (such as FedAvg) to aggregate all uploaded parameters.

[0105] Example: New global model W_global_new = 0.5 x ΔW1 + 0.5 x ΔW2; every month, the system uses knowledge distillation technology to "compress" the updated global model into a small model suitable for edge devices; Example: compress the original model from 10MB to less than 1MB, with a 3-fold increase in inference speed; the compressed model is pushed to each medical institution's blood cell separation device, replacing the old model and completing the upgrade; the entire process is periodically executed, ensuring that the local device always maintains the latest model capabilities and adapts to more clinical scenarios.

[0106] Step S150, based on the causal inference model, evaluate the potential risks of the pre-charge scheme, quantify the benefit-risk ratio of the scheme through counterfactual reasoning, and select the pre-charge scheme with the highest trust and risk indicators.

[0107] Among them, causal inference model (Causal Inference Model): a statistical modeling method for identifying causal relationships between variables, unlike correlation analysis, can simulate counterfactual scenarios of "what if different operations are taken". In this system, it is used to evaluate the impact of different priming schemes on patient safety. Counterfactual reasoning (Counterfactual Reasoning): In causal inference, by simulating alternative decisions to estimate the consequences that did not occur. For example: "If a different ratio of anticoagulants is used, will it lead to a higher risk of hemolysis?"

[0108] The system inputs the plurality of candidate priming schemes output by S130 into a causal inference model;

[0109] 1. Example candidate scheme: Scheme A: priming fluid = normal saline + ACD-A, priming volume = 60ml, anticoagulant ratio = 1:5; Scheme B: priming fluid = lactated Ringer's solution + ACD-B, priming volume = 70ml, anticoagulant ratio = 1:4. 2. Use causal diagrams and counterfactual reasoning to simulate changes in risk under different schemes; Example reasoning: "If scheme B is adopted, the expected hemolysis risk will increase by 1.5 percentage points"; Calculate the corresponding: expected hemolysis probability; expected coagulation risk; blood flow stability (from fluid mechanics prediction) for each scheme; Individualized adjustment combined with patient characteristics (such as body weight, underlying disease); Example: for patients with hypercoagulability, reduce the trust level of scheme B; 3. Calculate the benefit-risk ratio of each scheme using a weighted formula; Finally, select the priming scheme with the highest trust level and risk indicators that meet the standards; Example output: select scheme A, as it is the best in terms of safety and adaptability.

[0110] Adopting a model predictive control algorithm combined with real-time data and a dynamic model of the separation process, dynamically adjusting the centrifugal speed and acceleration parameters at a preset frequency, including:

[0111] Step S310, construct a multi-scale state space model containing blood rheology characteristics and centrifugal force field distribution, determine the preset state variables, generate a segmented linear model set through local linearization processing, and the preset state variables include hematocrit distribution and interface position.

[0112] Hemorheological Properties: Key parameters describing the behavior of blood flow, such as viscosity, red blood cell deformability, aggregation tendency, etc. Derived from laboratory test reports or real-time measurements by online viscosity sensors. Centrifugal Field Distribution: The distribution of centrifugal acceleration at different radii within the centrifuge chamber during the centrifugation process. Obtained through physical modeling or CFD simulation. State-Space Model: A mathematical representation used to describe the dynamic behavior of a system, typically composed of state equations and output equations. In this system, it is used to characterize the motion state of components during separation. Prescribed State Variables: Key parameters used to represent the operating state of the system, including: Hematocrit Distribution: Reflects the spatial distribution of blood components within the centrifuge chamber; Interface Position: The position of the boundary between, for example, the plasma and red blood cell layers; Derived from built-in optical / pressure sensors or image recognition modules within the device. Local Linearization: Approximation of a nonlinear system as a linear model near multiple operating points, facilitating the use of subsequent control algorithms. Completed by the internal mathematical engine of the system. Piecewise Linear Model Set: A collection of multiple local linear models covering the entire operating range, used to improve control accuracy.

[0113] The approximate process is described as follows (by example): Based on the current patient's blood sample data and device structure parameters, the system constructs a multi-scale state-space model that integrates hemorheological properties and centrifugal field distribution. Example input: hematocrit 42%, blood viscosity 3.8 mPa·s, rotation speed 3000 RPM; In this model, define key state variables: changes in red blood cell concentration over time and space; changes in the position of the plasma and red blood cell interface; Perform local linearization on the original nonlinear model to generate corresponding linear models at multiple typical operating points (such as different rotation speeds, different hematocrits); Example: generate linear model A at 3000 RPM, generate linear model B at 3500 RPM; Integrate these models into a piecewise linear model set for subsequent MPC controller use in different stages; The resulting model serves as the basis for the next step of rolling optimization and control strategy design.

[0114] Step S320, based on the finite time domain rolling optimization framework, taking the target component separation efficiency as the reference value, a cost function containing state error and control increment is constructed, and the sequence quadratic programming algorithm is used to solve the optimal control sequence of the future preset number of steps in real time.

[0115] Wherein, Model Predictive Control (MPC): an advanced control method based on dynamic model, which predicts the system behavior in the future and makes control decisions through online rolling optimization. Widely used in complex industrial process control. Finite Horizon Rolling Optimization Framework: the core mechanism of MPC, which only focuses on the optimization objective of a finite time step in the future in each control period, and constantly updates the prediction and control strategy. Target Component Separation Efficiency: the recovery rate of blood plasma or platelets, as the target reference value of MPC optimization. Derived from the preset process standard or doctor's set parameters. State Error: the deviation between the current state variable (such as interface position) and the expected value. Control Increment: the change amplitude of control input (such as speed change) between adjacent two control periods. Cost Function: a target function that measures the control performance, usually composed of state error and control increment. Sequential Quadratic Programming (SQP): a nonlinear optimization algorithm, suitable for solving the optimal control sequence with constraints in each control period. Provided by the built-in mathematical optimization module of the system.

[0116] The approximate process is described as follows (for example): 1. The system constructs the MPC cost function according to the current state (such as interface position offset 0.5 cm) and the preset target (such as stable at the center position); 2. Set the prediction horizon to 5 steps and the control horizon to 3 steps, which means that the system state evolution in the next 5 sampling periods will be considered each time; 3. Use the Sequential Quadratic Programming (SQP) algorithm to solve the cost function in real time to get a set of optimal future control sequence; example output: the next time speed + 50 RPM, the next time period + 30 RPM; 4. The controller only executes the first control action (such as immediately increasing the speed by 50 RPM), and then re-optimizes the prediction and optimization in the next period; the whole control process runs at a fixed frequency (such as once per second) to ensure that the separation efficiency is maximized and the operation is smooth.

[0117] Step S330, real-time identification of model parameters using recursive least squares algorithm, handling blood property time-varying influence through forgetting factor, triggering model structure adaptive adjustment when parameter change rate exceeds preset threshold.

[0118] Among them, the recursive least squares algorithm (Recursive Least Squares, RLS): an efficient algorithm for online estimation of system parameters, suitable for modeling tasks in dynamic changing environments. In this system, it is used to identify the changes of blood rheological properties over time. Forgetting factor: a key parameter in the RLS algorithm, used to weight new and old data, so that the model pays more attention to recent observations and weakens the influence of historical data, so as to better cope with time-varying characteristics. Time-varying blood properties: due to the mixing of anticoagulants, shear action or temperature changes during separation, blood viscosity, hematocrit and other parameters will change, which need to be compensated through real-time identification. Model parameter variation rate: represents the change amplitude of the current estimated parameter relative to the last time, which is used to determine whether to trigger model update. Threshold: if the parameter variation rate exceeds this value (for example, 5%), it is considered that the model is no longer applicable, and the structure needs to be rebuilt or adjusted. Model structure adaptive adjustment: when significant nonlinearity or model mismatch is detected, automatically switch to a more suitable model structure or re-localize linearization.

[0119] The approximate process is described as follows (by example): 1. The system continuously acquires state variables (such as interface position, speed, plasma flow) and control inputs during centrifugation; example inputs: current interface position = 1.2 cm; speed = 3200 RPM; plasma flow rate = 40 ml / min; 2. Use the recursive least squares algorithm (RLS) to identify the parameters of the current state space model in real time; example output: the current blood effective viscosity is estimated to be 4.1 mPa·s (the original model is 3.8 mPa·s); 3. Calculate the change rate of the current parameter and the last time; example calculation: (4.1-3.8) / 3.8≈7.9%; if the change rate exceeds the preset threshold (such as 5%), trigger the model structure adaptive adjustment mechanism; example action: switch from linear model A to linear model B to adapt to the kinetic changes caused by the increase of blood viscosity; the new model will be used as the prediction model of the next period MPC controller, improving control accuracy and robustness.

[0120] Step S340, introduce a polyhedral uncertainty set to describe model errors and measurement noise, convert the optimization problem into a robust constrained quadratic programming, and solve the control strategy that meets the preset safety constraint condition through scene optimization method.

[0121] Polyhedral Uncertainty Set: A mathematical structure used to describe system uncertainties, typically represented as a set of linear inequality constraints. Here it is used to model the range of influence of model mismatch and measurement noise. Model Mismatch: The deviation between the actual system behavior and the prediction model, possibly due to changes in blood rheological properties or local nonlinearities not accurately captured. Measurement Noise: Random errors from sensors (e.g. optical interface detector, pressure sensor) that affect the accuracy of state estimation. Its statistical properties are obtained through factory calibration or online filtering. Robust Constrained Quadratic Programming (RCQP): An optimization method that considers uncertainties, ensuring that safety constraints are still met under all possible perturbations. Scenario-based Optimization: A method that samples multiple "possible cases" from the uncertainty set, converting them into a set of deterministic constraints for joint optimization, reducing computational complexity. Safety Constraints: Include: maximum allowed rotational speed (e.g. ≤ 4000 RPM); upper limit on centrifugal acceleration (to avoid hemolysis); tolerance range for plasma layer position deviation (to ensure separation purity).

[0122] The system constructs a polyhedral uncertainty set based on the current model error and statistical information of sensor noise; example: assuming that the blood viscosity error does not exceed ±10%, the interface position measurement error is within ±0.2 cm; combine the MPC cost function with the safety constraints and the uncertainty set to convert it into a robust constrained quadratic programming problem (RCQP); example goal: minimize state error and control fluctuation for the next 5 steps; example constraints: even in the worst case, the rotational speed must not exceed 4000 RPM; use scenario optimization method to sample several "possible disturbance scenarios" from the uncertainty set; example: sample 5 different viscosity + measurement error combinations; apply the same control strategy to each scenario and verify whether all safety constraints are met; solve the optimal control input sequence under the premise of meeting all scenario constraints; example output: the optimal rotational speed increments for the next 3 steps are +60 RPM, +20 RPM, -10 RPM; the control system only executes the first step control action and renews the rolling optimization in the next period.

[0123] Step S350, construct a causal graph model of centrifugal parameters and separation effect, identify high-risk states based on pre-set risk quantification indicators, generate parameter correction schemes through counterfactual reasoning, and adjust the causal intervention of rotational speed and acceleration parameters.

[0124] Among them, the causal graph model: a modeling method that uses a directed graph to represent the causal relationship between variables, which is used to identify which input factors have a direct causal effect on the output result. In this system, it is used to analyze the causal relationship between centrifugal parameters and separation abnormalities. High-risk state: refers to the state that may occur adverse events under the current operating conditions, such as: red blood cell layer interface fluctuation too large → separation purity decreased; sudden change of rotation speed → hemolysis or pipeline pressure abnormal; derived from real-time monitoring data of equipment and clinical safety threshold. Risk quantification metric: such as hemolysis index rise rate, interface position deviation degree, etc., used to evaluate the risk level of the current state. Calculated by the built-in evaluation module of the system. Counterfactual reasoning: simulate the hypothetical reasoning of "whether the current risk can be avoided if a different control strategy is adopted" in the causal graph, which is used to generate correction schemes. Causal intervention adjustment: after identifying the high-risk state, intervene in the system behavior by changing the key control variables (such as reducing the acceleration change rate) to avoid risks.

[0125] The approximate process is described as follows (by way of example): 1. The system constructs a causal graph model based on real-time collected centrifugal process data (such as rotation speed, interface position, plasma flow rate), combined with historical operation records; 2. Example nodes: intervention variables: centrifugal rotation speed, acceleration; intermediate variables: blood shear force, red blood cell distribution uniformity; result variables: hemolysis index, separation efficiency; monitor each risk quantification metric in real time to determine whether to enter a high-risk state; example judgment rule: if the interface offset exceeds 0.5 cm and the rotation speed change rate > 200 RPM / s, it is marked as "high risk"; when a high-risk state is detected, the counterfactual reasoning mechanism is started; example reasoning: "if the acceleration is limited to within 100 RPM / s, can the current interface fluctuation be avoided?" Based on the results of counterfactual analysis, the system generates a parameter correction scheme; example output: adjust the acceleration of the next control step from +150 RPM / s to +80 RPM / s; at the same time, slightly increase the basic rotation speed to maintain the separation efficiency; the corrected control parameters are fed back to the MPC controller to participate in rolling optimization and execution in the next optimization period.

[0126] The separation process data is encrypted by a preset encryption algorithm and uploaded to the blockchain platform, and a maintenance plan is generated based on the abnormal detection results of the equipment operation data, including:

[0127] In step S510, feature engineering and privacy encryption are performed on the separation process data, and the device state fingerprint, operation log and consumable usage record are stored on the chain through the alliance chain platform.

[0128] Among them, the separation process data: equipment operating parameters (such as speed, pump speed), operation log, consumable record, etc. come from equipment sensors and control system. Feature engineering: extract key features from raw data, such as average value, volatility rate, state label, etc. Privacy encryption algorithm: such as AES-256, used for encryption processing of sensitive information (such as operator ID). Alliance chain platform: a blockchain system co-built by multiple trusted institutions, used for secure storage and sharing of medical data. Device state fingerprint: a set of simplified features reflecting the current state of the device, such as temperature, running time, load level.

[0129] Step S520, using graph neural network to analyze the causal relationship of abnormal events, and generating device abnormal root cause analysis report combined with blockchain storage data.

[0130] Graph Neural Network (GNN): a deep learning model specifically designed to handle graph structured data, capable of capturing complex relationships between nodes. In this system, it is used to analyze the causal relationship of abnormal events in device operation data.

[0131] Anomaly Event: refers to the abnormal phenomenon that occurs during the operation of the device, such as interface displacement too large, increased risk of hemolysis, etc. It comes from sensor monitoring and anomaly detection module.

[0132] Blockchain storage data (Blockchain Ledger Data): historical device state fingerprint, operation log and consumable usage record that has been chained, ensuring that the data is tamper-proof and transparent and traceable. Access through the alliance chain platform.

[0133] Root Cause Analysis Report: based on the analysis of abnormal events and their causal relationship, generate a detailed fault cause report to help technicians quickly locate the problem and take corrective measures.

[0134] The implementation process (example) is as follows: 1. The system reads the historical operation records of the device from the blockchain platform, including state fingerprint, operation log, consumable information; example input fields: device number SEP-202, timestamp, speed, pump speed, consumable batch, etc.; 2. Combine the currently detected abnormality (such as "interface position offset 0.6 cm"), build a causal graph containing device parameters, operation behavior and abnormal events; use graph neural network (GNN) to infer the causal influence between nodes, identify the main factors causing the abnormality; example output: the main cause of the abnormality is "motor vibration abnormality", which may be caused by "bearing wear"; generate root cause analysis report, including abnormality description, inferred cause, recommended measures.

[0135] Step S530, train the predictive maintenance model through federated learning, output the remaining useful life of the key components, and generate a maintenance work order containing spare parts information.

[0136] Wherein, Federated Learning: a distributed machine learning method that allows multiple participants to jointly train a model without sharing raw data. Suitable for privacy-sensitive fields such as medical devices. Predictive maintenance model: a model that predicts the remaining useful life (RUL) of key components by analyzing device operation data, allowing for early maintenance scheduling. Remaining useful life (RUL) of key components: based on the current state and historical data to predict the length of time the device's key components can still work normally. Maintenance work order: automatically generated work instructions based on prediction results, containing spare parts information, recommended maintenance time and steps.

[0137] The implementation process (example) is as follows: 1. The system collects device operation data from multiple medical institutions, which are retained locally and jointly trained through a federated learning framework; example input fields: speed, temperature, pressure, operation log, etc.; 2. Use federated learning to train a predictive maintenance model that can predict the remaining useful life of the device's key components based on real-time data; example model output: centrifuge bearing is expected to run for another 500 hours; 3. Based on the prediction results, the system automatically generates detailed maintenance work orders.

[0138] The maintenance work order not only includes the necessary spare parts information, but also provides detailed operation guidelines to ensure that maintenance personnel can efficiently complete the task; the work order is pushed to the relevant maintenance team through the system and recorded on the blockchain platform for subsequent audit use.

[0139] Step S540, deploy smart contracts, automatically trigger maintenance processes based on device runtime, predicted lifespan, or repeated abnormalities, including device locking, work order distribution, and spare parts pre-allocation.

[0140] Wherein, Smart Contract: an automatically executable program deployed on the blockchain, which triggers corresponding operations such as generating work orders and locking devices based on preset conditions. Device runtime: cumulative device working time used to determine if it has reached the regular maintenance period. Derived from the device's internal timer. Predicted lifespan: the remaining useful life (RUL) of key components output by the predictive maintenance model, used to schedule replacement in advance. Repeated abnormalities: the same type of abnormality occurring multiple times within a set time, which may indicate potential failure risks. Derived from the system's abnormality recording module.

[0141] The implementation process (for example) is as follows: 1. The system continuously monitors the equipment status, including the running time, the predicted life of key components, and the frequency of abnormal events; example input: bearing remaining life = 80 hours, interface deviation abnormality occurred 2 times in the last 3 days; 2. When any of the following conditions are met, trigger the smart contract to execute: the equipment running time reaches the preset maintenance period (such as every 1000 hours); the predicted life of a component is below the threshold (such as <100 hours); the same abnormality occurs repeatedly within a set window period (such as ≥2 times in 3 days); 3. After triggering, the smart contract automatically performs the following operations: equipment locking: restricts the device from continuing to use, preventing it from running with defects; work order distribution: pushes the maintenance work order to designated maintenance personnel or service center; spare parts pre-allocation: according to the work order content, reserve the required spare parts from the inventory system; all operations are recorded on the chain for evidence, ensuring that the process is transparent and traceable; example record fields: trigger time, trigger reason, processing personnel, completion status.

[0142] Step S550, after maintenance, perform function, performance and clinical effect self-check, the results are recorded on the chain for evidence and used to optimize the predictive maintenance model.

[0143] Among them, function self-check: a series of detection procedures that the device automatically runs after maintenance is completed, verifying whether the core functions have returned to normal. Such as motor response, sensor accuracy, etc. Performance self-check: evaluate whether the key performance indicators of the device meet the standards, such as separation efficiency, speed control accuracy, etc. Clinical effect self-check: assess the impact of the device on blood component separation quality through simulation or actual operation, such as hemolysis index, recovery rate, etc. Predictive maintenance model: a machine learning model used to predict the remaining life of device components and generate maintenance plans, trained by federated learning. Model feedback loop: feedback maintenance and operation data to the model training process to improve future prediction accuracy.

[0144] The implementation process (for example) is as follows: 1. After maintenance is completed, the system automatically starts the function, performance and clinical effect self-check process; example detection items: function: centrifuge start-stop response is normal; performance: speed control error <±20 RPM; clinical: simulated separation sample hemolysis index ≤0.1 g / dL; 2. All detection results generate structured reports and are uploaded to the blockchain platform for evidence; example fields: detection time, detection type, result value, whether passed; if the detection passes, the device is unlocked and normal use is resumed; 3. At the same time, this maintenance and detection data is fed back as feedback information to the retraining process of the predictive maintenance model; example feedback content: "vibration is reduced by 30% after bearing replacement, RUL is reset to the initial value"; after the model is updated regularly, it is distributed to each device end, forming a continuous optimization closed loop.

[0145] The results are recorded on the chain for evidence and used to optimize the predictive maintenance model, including:

[0146] Step S551, after maintenance, perform function, performance and clinical effect self-check, feature extraction and standardization of self-check result data, form training samples in unified format.

[0147] Feature extraction (Feature Extraction): Extract key statistics or state description from original detection data for modeling analysis. Standardization (Standardization): Unify data from different sources and units to the same scale for model training.

[0148] The implementation process (example) is as follows: 1. After maintenance, the system performs self-check and obtains device state data; Example: speed error = +18 RPM, hemolysis index = 0.09 g / dL; 2. Extract key features and perform standardization; Example feature vector: [speed error, hemolysis index change rate]→ standardized to [-0.35, 0.12]; 3. Packaged as training samples in unified format for model update.

[0149] Step S552, through the federated learning framework, combine the local self-check result data with the data of multiple medical institutions for joint training, and update the parameters of the predictive maintenance model.

[0150] Federated learning (Federated Learning): A distributed machine learning method that allows multiple participants to jointly train a model without sharing raw data. Self-check result data: Function, performance and clinical effect detection data generated by self-check after device maintenance. Training samples: Structured data after feature extraction and standardization, used for model training. Predictive maintenance model: A model used to predict the remaining useful life of key components of a device and generate maintenance plans.

[0151] The implementation process is as follows: 1. Each medical institution extracts and standardizes the self-check result data after local maintenance to form training samples in unified format; these training samples are uploaded to the central server through the federated learning framework, but do not include raw data to ensure data privacy; example upload content only contains model parameter update information; 2. The central server aggregates the model parameter updates uploaded by each institution and generates a global updated model using weighted average method; example formula: new global model = Σ (each institution model update * weight) / total weight; 3. The updated predictive maintenance model is distributed back to each participating institution to replace the old version of the model; example operation: download and deploy the new model to the local device control system; the model is updated regularly to continuously optimize to improve prediction accuracy and reliability.

[0152] Step S553, build a reward function centered on maintenance effect, use deep Q network reinforcement learning framework, trigger parameter adjustment optimization after each maintenance.

[0153] Reward function: a numerical function used to measure the quality of system state, designed with maintenance effectiveness as the core indicator. Deep Q Network: a reinforcement learning algorithm that combines Q-learning with deep neural networks, used to automatically learn optimal control strategies in complex environments. Post-maintenance parameter adjustment: dynamically optimizing key parameters in the predictive maintenance model based on device self-test results.

[0154] The implementation process is as follows: 1. After each device completes maintenance and generates a self-test report, the system extracts relevant features as reinforcement learning input state; example state: [current speed error, hemolysis index change rate, maintenance type]; 2. Construct a reward function with maintenance effectiveness as the core, score based on self-test pass and device running stability; example rule: if self-test passes and runs smoothly → reward value +1; if alarm again → reward value -0.5; 3. Use the Deep Q Network (DQN) framework to train the agent, learn what parameter adjustment can achieve higher rewards under different states; example action: adjust the life decay coefficient of a component in the predictive model; 4. Trigger parameter optimization adjustment once after each maintenance, update the behavior strategy of the predictive maintenance model; example output: improve the identification sensitivity of bearing wear type faults.

[0155] The reward function with maintenance effectiveness as the core is constructed as follows:

[0156] R = 0.6 x recovery rate + 0.3 x (-Δ hemolysis rate) + 0.1 x (-Δ maintenance cost).

[0157] Where Δ is the change in the index before and after maintenance, the Deep Q Network (DQN) is used as the reinforcement learning framework, the state space is the real-time feature vector of the device, and the action space is the parameter adjustment strategy of the predictive model (such as adjusting the number of LSTM hidden layer neurons ±10%). Trigger 100 steps of reinforcement learning training after each maintenance, optimize the model decision logic through the experience replay pool (capacity 100,000), so that the predictive maintenance strategy is more matched with the actual clinical effectiveness.

[0158] Step S554, combine historical maintenance records to construct a multi-objective optimization function, and use the NSGA-II algorithm to dynamically optimize the prediction logic.

[0159] Where historical maintenance records: device past maintenance time, type, replacement parts, self-test results, etc. information, from system logs and blockchain storage data. Multi-objective optimization function: optimize multiple conflicting or related objectives at the same time, such as prediction accuracy, maintenance cost, device availability, etc. NSGA-II algorithm: an efficient multi-objective genetic optimization algorithm used to find the Pareto optimal solution set in complex solution space.

[0160] Assume target 1 as follows: minimize prediction error, expressed as follows: Where T i is the actual lifetime, target 2 as follows: minimize false positive rate, example expression as follows: Target 3 as follows: maximize device availability, example expression as follows: In summary, the above three objectives are unified into a multi-objective vector function: min F(x) = [f1(x), f2(x), -f3(x)]. x is the model adjustable parameter vector, such as feature weight, decay coefficient, threshold offset, etc.

[0161] The NSGA-II genetic algorithm is used to solve the above multi-objective function, the process is as follows: 1, initialization of population: randomly generate multiple model parameter combinations; 2, evaluate fitness: calculate the objective function value of each individual; 3, non-dominated sorting: sort the individuals according to the Pareto frontier; 4, crossover and mutation: generate new individuals and update the population; 5, termination condition: reach the maximum iteration number or converge stably; 6, output the Pareto optimal solution set: for the system to choose the most suitable strategy deployment.

[0162] Step S555, performance verification of the optimized predictive maintenance model, using the reserved validation dataset to evaluate the prediction accuracy and reliability of the model, and updating the model according to the verification results.

[0163] Among them, the reserved validation dataset: a part of data deliberately left out during the training process, used to independently evaluate the performance of the model, to ensure the generalization ability of the model. Prediction accuracy: a measure of the accuracy of the model's prediction of future events, usually expressed in terms of error or accuracy. Reliability: refers to the ability of the model to maintain consistent performance under different conditions, including stability and robustness. Model iterative update: the process of adjusting and optimizing the model according to the verification results to improve its long-term performance.

[0164] The results are stored on the chain and used to optimize the predictive maintenance model, including:

[0165] Step Sa00, based on the digital twin module, combining the physical model of the device and historical failure data, generating a preset number of virtual failure samples covering single, complex and extreme working conditions.

[0166] The digital twin module uses multi-physical field coupling modeling method, integrates fluid mechanics, thermodynamics and biomechanics model, and constructs a 1:1 mapping three-dimensional virtual model with the actual device; the reinforcement learning agent uses the deep deterministic policy gradient (DDPG) algorithm to train the device life cycle maintenance benefit maximization target.

[0167] In the sample generation process, first, based on the historical fault database, the characteristic parameters of common failure modes such as bearing wear and sensor drift are extracted as the basic template; second, using the generative adversarial network (GAN) under the constraint of medical knowledge graph, random generation of parameter combination of composite failure (such as “bearing wear + flow sensor failure”) and extreme working condition (such as parameter disorder after power failure and restart); finally, input the above fault parameters into the digital twin model, and generate virtual fault samples containing multi-dimensional data such as vibration spectrum and pressure fluctuation curve through simulation deduction. The number of generated samples is dynamically determined by the preset fault coverage threshold (such as covering 95% of historical fault types), ensuring the diversity of virtual scenarios and the consistency with clinical actual risks.

[0168] Step Sb00, mix the virtual fault data and the actual maintenance data stored in the blockchain according to the preset proportion to construct a dynamically updated training data set.

[0169] Virtual fault data includes device operating state data, fault feature parameters and simulation results generated by the digital twin module, covering multi-modal information such as pressure, flow, vibration, etc.; the actual maintenance data stored in the blockchain is the real operating data such as device operation log, abnormal detection record and maintenance work order after being processed by a preset encryption algorithm and uploaded to the chain, which has the characteristics of non-tamperability and traceability.

[0170] The preset proportion is dynamically adjusted according to data reliability evaluation: when the accuracy rate of virtual fault data verified by clinical verification is more than 90%, the proportion can be increased to 60%; otherwise, it is reduced to 40%, ensuring that the proportion of actual maintenance data is not less than 60%. Using data stratified sampling method, classify and match the two types of data according to time sequence and fault type, for example, for “centrifugal motor fault” category, extract a preset number of samples from virtual and actual data respectively.

[0171] Build a dynamic updating mechanism: every 10 actual maintenance operations, automatically trigger the data set updating process, replace the obsolete samples in the original data set with the newly generated maintenance data after feature extraction and standardization processing, and update the data labels at the same time, ensuring that the training data set always fits the latest operating state of the equipment, providing high-quality training data support for the predictive maintenance model.

[0172] Step Sc00, extract the equipment performance degradation trend predicted by the digital twin model, and convert it into a dynamic constraint condition of the multi-objective optimization function.

[0173] Key performance parameter degradation features are extracted from the time series data output by the digital twin model, including centrifugal motor residual life prediction value \(\hat{L}\), bearing fault false alarm frequency F, and device downtime ratio D. Convolutional neural networks (CNN) are used to extract features from multi-modal data such as vibration frequency spectrum and pressure fluctuation. The attention mechanism is used to identify parameters with high correlation to the objective function (such as the correlation weight between motor efficiency decay rate and prediction error, which reaches 0.7).

[0174] The three optimization objectives are converted into dynamic constraint conditions:

[0175] Prediction error constraint: based on the minimum prediction error function of target 1 Set constraints Where σ1 = 0.8 is the historical error mean scaling coefficient. When the digital twin predicted degradation trend exceeds 1.5 times the historical mean, it is automatically tightened to σ1 = 0.6.

[0176] False alarm rate constraint: according to the false alarm rate function of target 2 Set constraint F rate ≤ 5%, if the false alarm rate of the composite fault scenario in the virtual simulation exceeds this threshold, trigger the weight adjustment mechanism, increase the false alarm rate target weight from 0.3 to 0.4.

[0177] Device availability constraint: based on the availability function of target 3 Set constraint A ≥ 95%, when the digital twin predicted maintenance downtime may cause a decrease in availability, use the NSGA-II algorithm to prefer strategies with tighter maintenance windows.

[0178] Embed the above constraints into the multi-objective vector function Where θ is the model adjustable parameter. Use the NSGA-II algorithm with constraint processing. In the genetic operation, introduce the feasibility rule: for individuals that violate the constraints, sort them according to their constraint violation degree; when the digital twin predicted degradation trend changes abruptly (such as a 20% increase in vibration frequency), automatically trigger the elite preservation strategy, migrate the constraint threshold of the historical optimal solution to the current optimization process, and ensure the real-time adaptability of the maintenance strategy.

[0179] Step Sd00, introduce a dynamic weight coefficient for clinical risk, use the improved NSGA-III algorithm to solve the Pareto front solution set, and optimize the maintenance strategy.

[0180] Dynamic weight construction: based on the digital twin simulation of hemolysis probability, coagulation risk, and other clinical indicators, combined with patient-specific parameters (such as platelet count) and device status, dynamically adjust the risk indicator weight (such as the hemolysis risk weight floating ±20% in real time).

[0181] Objective function extension: Clinical risk indicators are used as the fourth dimension objective, forming a multi-objective vector function with prediction error, false positive rate, and device availability. The NSGA-II algorithm with constraint handling is used to solve it.

[0182] Algorithm improvement and optimization: Adaptive crossover and mutation operators are introduced to dynamically adjust genetic parameters based on iteration number. The hyper-volume indicator is used to screen the Pareto front, ensuring the diversity of solutions within the clinical safety threshold (e.g., hemolysis probability ≤ 5%).

[0183] Strategy generation and application: The strategy with the highest comprehensive score is selected from the Pareto solution set, automatically converted into maintenance parameters (e.g., component replacement priority), and the strategy effectiveness is verified through blockchain data storage.

[0184] Step Se00, simulate a preset number of stress test scenarios in a virtual environment to verify the reliability of the predictive maintenance model.

[0185] Based on the digital twin model, a stress test scenario library is constructed, covering extreme working conditions (e.g., motor overload 200%), composite failures (bearing wear + sensor failure), and rare events (continuous power failure restart). Using the generative adversarial network (GAN), a preset number of abnormal parameter combinations are automatically generated, such as setting the centrifugal speed to 1.5 times the rated value and superimposing ±10% fluctuation. Run the maintenance strategy through virtual simulation, monitor the response results of the predictive maintenance model, and verify whether its core indicators such as fault identification accuracy and maintenance timing prediction error meet the threshold requirements (e.g., false positive rate ≤ 3%), ensuring the reliability of the model in complex scenarios.

[0186] Step Sf00, compare virtual prediction with actual operation data, and when the deviation exceeds the preset threshold, use the structural causal model to trace the root cause.

[0187] Compare the device performance parameters predicted by the digital twin model (e.g., centrifugal motor speed, pressure fluctuation curve) with the actual sensor data in real time, and calculate the prediction error (e.g., remaining life prediction deviation > 15%). When the deviation exceeds the preset threshold, based on the structural causal model in Step Sf00, divide the intervention variables (e.g., maintenance strategy adjustment), outcome variables (device failure), and covariates (patient blood characteristics, environmental temperature), eliminate the confounding factors through the backdoor criterion. Use the do-operator to simulate intervention, combine Bayesian network to calculate the causal effect value of each factor on the deviation, locate the key factors causing the deviation (e.g., sensor aging, algorithm parameter mismatch), and generate a root cause analysis report.

[0188] Step Sg00, based on the root cause analysis results, automatically adjust the model parameters and optimization function weights, trigger model reconstruction and strategy iteration, and form a closed-loop optimization.

[0189] Based on the root cause analysis report of step Sf00, if the positioning is algorithm parameter mismatch, the LSTM network weight, feature extraction threshold and other parameters in the predictive maintenance model are automatically adjusted;If it is determined that sensor aging causes data deviation, the filter coefficient of the data preprocessing module is corrected.Meanwhile, according to the root cause type, the weight of each target in the multi-objective optimization function is dynamically adjusted, such as the deviation caused by the misjudgment of clinical risk, the weight of clinical risk is increased by 0.2.The model reconstruction program is triggered, the model is retrained by using the latest virtual simulation data and the actual maintenance data stored in the block chain, and the optimized maintenance strategy is pushed to the equipment for execution, completing the closed loop process of "analysis-adjustment-optimization-application".

[0190] Based on the same inventive concept, the embodiments of the present application provide an intelligent management and control system for a blood cell separation machine, which comprises a memory and a processor, and the memory stores a program capable of running on the processor to realize the method of any one of the embodiments of the present application. Figures 1 to 2 The program of any one method.

Claims

1. A blood cell separator intelligent management and control method, characterized in that, The method comprises the following steps: starting a blood cell separator and performing a sensor self-checking program, importing a patient blood test report and physiological parameters, matching an adaptive basic separation scheme from a plurality of preset separation schemes through a preset classification algorithm, and automatically configuring a disposable pipeline pre-charging scheme including a pre-charging fluid type, a pre-charging amount and an anticoagulant ratio based on the patient physiological parameters; after completing the pipeline pre-charging based on the pre-charging scheme, using a multi-modal sensing unit to collect pressure, flow, optical and vibration data in real time at a preset sampling frequency, and constructing a separation process dynamic model including a preset number of state variables after preprocessing using a data fusion algorithm; obtaining patient blood indicators, inputting the patient blood indicators into a preset deep neural network, selecting an adaptive scheme from a plurality of preset sub-separation modes and generating an initial centrifugal parameter combination, and using a model predictive control algorithm to combine the real-time collected data and the separation process dynamic model to dynamically adjust the centrifugal speed and acceleration parameters at a preset frequency; real-time monitoring of the state variables of the separation process dynamic model and the sensor data, and starting the parameter compensation mechanism corresponding to the abnormal type when the monitoring parameters deviate from the preset range; uploading the separation process data to a blockchain platform after encryption through a preset encryption algorithm, and generating a maintenance plan based on the abnormal detection result of the equipment operation data.

2. The intelligent management and control method for blood cell separator according to claim 1, characterized in that, The method of matching an adaptive basic separation scheme from a plurality of preset separation schemes through a preset classification algorithm comprises: mapping the numerical features in the patient blood test report to a semantic vector space through contrastive learning, and converting the physiological parameters into high-dimensional embedding vectors using a medical knowledge graph, and forming a patient feature tensor containing clinical semantics through an attention mechanism; based on a model-independent meta-learning framework, using a pre-trained medical field meta-model to perform gradient updates on the patient feature tensor for a preset number of times to generate a special classifier for the current patient; constructing a fuzzy reasoning engine through a preset number of clinical rules, combining the output results of the special classifier, the time series analysis model and the image recognition model through Dempster-Shafer evidence theory, and calculating the trust score of each basic separation scheme; aggregating the classification model parameters of multiple medical institutions through a federated learning framework to periodically update the global classifier; evaluating the potential risks of each candidate scheme based on a causal reasoning model, quantifying the benefit-risk ratio of the scheme through counterfactual reasoning, and selecting the basic separation scheme with the highest trust score and meeting the risk indicators, wherein the potential risks include hemolysis probability and coagulation risk.

3. The intelligent management and control method of a blood cell separator according to claim 2, wherein, The method of evaluating the potential risks of each candidate scheme based on a causal reasoning model, quantifying the benefit-risk ratio of the scheme through counterfactual reasoning, and selecting the basic separation scheme with the highest trust score and meeting the risk indicators comprises: constructing a structural causal model including patient features, blood indicators and device parameters, dividing intervention variables, outcome variables and covariates, and identifying key influencing factors through backdoor criteria, wherein the intervention variables are separation schemes, the outcome variables include hemolysis and coagulation risks, and the covariates include clinical confounding factors; simulating intervention using a do-operator, and calculating the causal effect value of each scheme on risk through a Bayesian structural equation model; Estimate potential traits based on patient data, calculate risk differences of different schemes through counterfactual queries; Adopt Dempster-Shafer theory to fuse causal effects, fuzzy reasoning confidence, and historical case data according to preset weights, construct a three-dimensional evaluation matrix to calculate a comprehensive score; According to the preset safety threshold and benefit threshold, filter the schemes that meet the standard and have the optimal risk-benefit ratio.

4. The intelligent management and control method of a blood cell separator according to claim 3, characterized in that, According to the preset safety threshold and benefit threshold, filter the schemes that meet the standard and have the optimal risk-benefit ratio includes: Set the safety threshold: hemolysis risk probability < 5%, coagulation risk probability < 8%, and set the benefit threshold: separation efficiency > 90%, recovery rate > 85%; Iterate through all candidate schemes, filter the feasible solution set that meets both the safety threshold and the benefit threshold; Perform multi-objective optimization analysis on the schemes in the feasible solution set, identify the Pareto optimal solution set; Calculate the risk-benefit ratio for each scheme in the Pareto optimal solution set, and the specific risk-benefit ratio calculation formula is as follows: wherein w k is the benefit index weight, v l is the risk index weight; Adjust threshold weight according to clinical scenario, and select RRR i The largest solution as the final output.

5. The intelligent management and control method for blood cell separator according to claim 1, characterized in that, Simultaneously configure the disposable tubing pre-charge scheme including pre-charge fluid type, pre-charge volume, and anticoagulant ratio based on patient physiological parameters includes: Construct a causal knowledge graph containing patient characteristics, blood indicators, and pre-charge parameters, and use graph neural networks to generate patient feature tensors; Based on the model-independent meta-learning framework, use pre-trained meta-models to perform gradient updates on patient feature tensors for a preset number of times to generate specialized configuration models; Build a fuzzy reasoning engine through a preset number of clinical rules, and combine the output results of the specialized configuration model, hemodynamic prediction model, and equipment compatibility data to dynamically adjust the pre-charge parameters; Aggregate model parameters from multiple medical institutions through a federated learning framework, periodically update the global model, and use knowledge distillation techniques to compress it to edge devices; Evaluate the potential risks of the pre-charge scheme based on the causal reasoning model, quantify the benefit-risk ratio of the scheme through counterfactual reasoning, and select the pre-charge scheme with the highest confidence and risk indicators that meet the standard.

6. The intelligent management and control method for blood cell separator according to claim 1, wherein, Use model predictive control algorithm combined with real-time data and separation process dynamic model to dynamically adjust centrifugal speed and acceleration parameters at preset frequency includes: Construct a multi-scale state space model containing blood rheological properties and centrifugal force field distribution, determine the preset state variables, and generate a segmented linear model set through local linearization processing, the preset state variables include hematocrit distribution and interface position; Based on the finite time domain rolling optimization framework, take the target component separation efficiency as the reference value, construct a cost function containing state error and control increment, and use the sequential quadratic programming algorithm to solve the optimal control sequence for the future preset number of steps in real time; Use recursive least squares algorithm to identify model parameters in real time, and use forgetting factor to handle the time-varying influence of blood properties, trigger model structure adaptive adjustment when the parameter change rate exceeds the preset threshold; Introduce a polyhedral uncertainty set to describe model errors and measurement noise, convert the optimization problem into a robust constrained quadratic programming problem, and solve the control strategy that meets the preset safety constraints through scenario optimization method; A causal diagram model of centrifugation parameters and separation effect is constructed, a high-risk state is identified based on a preset risk quantification index, a parameter correction scheme is generated through counterfactual reasoning, and the speed and acceleration parameters are adjusted through causal intervention.

7. The intelligent management and control method for blood cell separation machine according to any one of claims 1 to 6, characterized in that, The separation process data is encrypted through a preset encryption algorithm and uploaded to the blockchain platform, and a maintenance plan is generated based on the abnormal detection results of the equipment operation data, including: Feature engineering and privacy encryption are performed on the separation process data, and device state fingerprints, operation logs, and consumable usage records are stored on the blockchain platform; Graph neural networks are used to analyze the causal relationships of abnormal events, and a device abnormal root cause analysis report is generated based on the blockchain stored data; A predictive maintenance model is trained through federated learning, which outputs the remaining useful life of key components and generates a maintenance work order containing spare parts information; An intelligent contract is deployed to automatically trigger maintenance processes, including device locking, work order distribution, and spare parts pre-allocation, based on equipment runtime, predicted lifespan, or repeated abnormalities; After maintenance, perform a self-check of functionality, performance, and clinical effectiveness, and store the results on the blockchain for optimizing the predictive maintenance model.

8. The intelligent management and control method of a blood cell separator according to claim 7, wherein, The results are stored on the blockchain and used to optimize the predictive maintenance model, including: After maintenance, perform a self-check of functionality, performance, and clinical effectiveness, and store the results on the blockchain for optimizing the predictive maintenance model, including: Feature extraction and standardization of self-check result data are performed to form a unified format training sample; Through the federated learning framework, the local self-check result data is jointly trained with the data of multiple medical institutions to update the parameters of the predictive maintenance model; A reward function is constructed with maintenance effect as the core, and a deep Q network reinforcement learning framework is used to trigger parameter adjustment and optimization after each maintenance; A multi-objective optimization function is constructed based on historical maintenance records, and the NSGA-II algorithm is used to dynamically optimize the prediction logic; 9. A blood cell separator intelligent management and control system, characterized in that, The performance of the optimized predictive maintenance model is verified, and the reserved validation dataset is used to evaluate the prediction accuracy and reliability of the model. According to the validation results, the model is iteratively updated. A blood cell separation machine intelligent management and control method as claimed in any one of claims 1 to 8 is implemented when the program is loaded and executed by the processor.

Images