Automatic control method and system for field intelligent injection pump based on physiological parameter monitoring

Abstract

The application provides an automatic control method and system for a field intelligent injection pump based on physiological parameter monitoring, and relates to the technical field of automatic control, which comprises the following steps: collecting physiological parameter information through a flexible electronic patch; inputting the physiological parameter information into an intelligent decision module, generating an individualized drug administration scheme based on a reinforcement learning algorithm in combination with a physiological function state report and a drug action mechanism knowledge base, wherein the individualized drug administration scheme comprises drug proportioning, injection rate and injection timing; sending the individualized drug administration scheme to an intelligent injection module to execute an injection step; an energy management unit collects and stores energy through a flexible photovoltaic film and a phase change energy storage material, establishes a multi-objective optimization function based on drug storage conditions and environmental parameters, intelligently distributes and adaptively adjusts the power consumption of the injection pump system; evaluates the state of the injection pump pipeline based on multi-modal sensing information, dynamically evaluates safety in combination with real-time physiological monitoring, and triggers hierarchical protection.

Description

Technical Field

[0001] This invention relates to automatic control technology, and more particularly to an automatic control method and system for a field intelligent infusion pump based on physiological parameter monitoring. Background Technology

[0002] In field environments, intelligent infusion pump systems play a crucial role in patient treatment. Traditional infusion pump systems rely primarily on healthcare professionals' experience and fixed dosages for drug administration, making precise adjustments based on the patient's real-time physiological state difficult. Existing intelligent infusion pump systems for field use still have several shortcomings. First, most systems lack adaptability to complex field environments, failing to compensate for injection parameters in real time based on environmental factors such as temperature and humidity, thus affecting drug delivery accuracy. Second, the energy management of existing systems is relatively simple, making it difficult to achieve continuous operation for extended periods in field environments, limiting the system's practicality. Finally, most systems lack dynamic safety assessment mechanisms for the injection process, failing to promptly detect and address potential risks such as infusion line blockage, posing safety hazards. Summary of the Invention

[0003] This invention provides an automatic control method and system for intelligent infusion pumps in the field based on physiological parameter monitoring, which can solve the problems in the prior art.

[0004] A first aspect of the present invention,

[0005] A method for automatic control of intelligent infusion pumps in the field based on physiological parameter monitoring is provided, including:

[0006] The patient's heart rate, respiration, blood pressure, blood oxygen saturation and body surface electrical signals are collected by a flexible electronic patch. The flexible electronic patch has a built-in signal processing chip to perform signal conditioning and noise reduction on the collected information to obtain physiological parameter information.

[0007] Physiological parameter information is input into an intelligent decision-making module, which includes a physiological parameter analysis unit and a medication optimization unit. The physiological parameter analysis unit performs multimodal feature extraction and trend prediction on the physiological parameter information to generate a physiological function status report. The medication optimization unit generates a personalized dosing plan based on a reinforcement learning algorithm, combined with the physiological function status report and a drug action mechanism knowledge base. The personalized dosing plan includes drug ratio, injection rate, and injection sequence.

[0008] The personalized drug delivery plan is sent to the intelligent injection module, which includes a micro-peristaltic pump mechanism, a multi-sensor array, a control processor, and an energy management unit. The control processor adaptively compensates for the injection parameters based on the ambient temperature, humidity, and air pressure parameters collected in real time by the multi-sensor array, and executes the compensated injection parameters through the micro-peristaltic pump mechanism. The energy management unit collects and stores energy through flexible photovoltaic thin films and phase change energy storage materials, and establishes a multi-objective optimization function based on drug storage conditions and environmental parameters to intelligently allocate and adaptively adjust the power consumption of the injection pump system. The system evaluates the status of the injection pump tubing based on multimodal sensor information, performs dynamic safety assessment in conjunction with real-time physiological monitoring, and triggers graded protection.

[0009] In one alternative implementation,

[0010] The steps for generating a physiological function status report by performing multimodal feature extraction and trend prediction on the physiological parameter information through the physiological parameter analysis unit include:

[0011] Multimodal feature extraction is performed on the physiological parameter information. Temporal morphological features are extracted using a deep convolutional network, and frequency domain features are extracted using a fast Fourier transform and a spectral encoder. Nonlinear dynamic features are extracted by calculating sample entropy and Lyapunov exponent. The deep convolutional network uses batch normalization and ReLU activation function, and the spectral encoder uses layer normalization. A bidirectional attention-enhanced long short-term memory network is used to model the temporal dependency of the multimodal features. The long short-term memory network includes a bidirectional long short-term memory layer for forward and backward information fusion, a multi-head self-attention layer for establishing temporal correlation weights, and a residual connection layer for preserving the original features.

[0012] A multi-scale trend prediction framework is constructed based on the results of temporal dependency modeling, including a short-term predictor based on gated recurrent unit network for five-minute scale prediction, a medium-term predictor based on Transformer encoder for thirty-minute scale prediction, and a long-term predictor based on temporal convolutional network for four-hour scale prediction.

[0013] The multi-scale trend prediction results are input into the cardiovascular system scoring unit and the respiratory system scoring unit based on the gated attention mechanism to score physiological indicators. The cardiovascular system scoring unit calculates the score based on heart rate variability, blood pressure fluctuation characteristics and body surface electrical signal morphology characteristics. The respiratory system scoring unit calculates the score based on respiratory rate changes, blood oxygen saturation trends and respiratory rhythm characteristics.

[0014] Trend analysis was performed on the physiological index scores. The trend analysis used an ensemble learning strategy to fuse the prediction results of the autoregressive moving average model and the Prophet model, and the confidence interval of the prediction results was calculated using the Monte Carlo sampling method.

[0015] A physiological function status report is generated based on the confidence intervals of the physiological indicator scores, trend analysis, and prediction results. The physiological function status report includes cardiovascular system stability analysis results, respiratory system dynamic characteristics, autonomic nervous system regulation capacity indicators, organ function status quantification results, and physiological rhythm homeostasis evaluation.

[0016] In one alternative implementation,

[0017] The medication optimization unit generates a personalized dosing regimen based on a reinforcement learning algorithm combined with the physiological function status report and the drug action mechanism knowledge base. The personalized dosing regimen includes steps such as drug ratio, injection rate, and injection sequence:

[0018] A drug action mechanism knowledge base is constructed. Drug molecular structures are encoded using an attention-enhanced graph neural network to obtain structured representation vectors. Transduction learning is then used to transfer knowledge of known drug action mechanisms to new drugs, forming an entity relationship network containing information on drug molecular structures, targets, metabolic pathways, and drug interactions. Indicators in physiological function status reports are standardized and dimensionality-reduced preprocessing is performed. Historical dosing records and physiological indicator predictions are combined to form a dosing environment state vector. An action space is designed, comprising discretizing drug combinations into standard ratio schemes to form a drug ratio space, discretizing injection rate ranges into multiple levels to form an injection rate space, and dividing dosing time into time segments to form an injection time sequence space.

[0019] By combining the drug action mechanism knowledge base and the drug administration environment state vector, two identical but parameter-independent dual deep Q-networks are constructed. One network is used to select the dosing regimen, and the other network is used to evaluate the Q value of the selected dosing regimen. A comprehensive reward function is designed, which includes physiological indicator improvement rewards, drug interaction penalties, and dosing safety constraints. The weight coefficients of the comprehensive reward function are determined by grid search optimization. An initial dosing regimen is generated based on the trained deep Q-network. A regimen safety assurance mechanism is constructed. The contribution of indicators in the physiological function status report, historical dosing records, and predicted changes in physiological indicators to the Q value is analyzed using SHAP values. The safety of the initial dosing regimen is verified based on the Monte Carlo method. The risk of the initial dosing regimen is assessed by sampling the state transition probability. When a potential risk is detected, an early warning mechanism is activated and the parameters of the initial dosing regimen are automatically adjusted.

[0020] Collect physiological response data to update the drug administration environment state vector, calculate the actual comprehensive reward function value and store it in the experience replay pool, sample data from the experience replay pool to update the strategy network, and when a change in physiological state is detected to exceed a preset change threshold, trigger the drug administration plan adjustment and re-execute the plan safety assurance mechanism to generate the final personalized drug administration plan.

[0021] In one alternative implementation,

[0022] By combining the drug action mechanism knowledge base and the drug administration environment state vector, two identical but parameter-independent dual deep Q-networks are constructed. One network is used to select the dosing regimen, and the other network is used to evaluate the Q-value of the selected regimen. A comprehensive reward function is designed, including rewards for improving physiological indicators, penalties for drug interactions, and dosing safety constraints. The weight coefficients of the comprehensive reward function are determined through grid search optimization. The steps for generating an initial dosing regimen based on the trained deep Q-network include:

[0023] Two identical but parameter-independent dual-depth Q-networks are constructed. Each dual-depth Q-network includes an input layer, three fully connected layers, a batch normalization layer, and a random deactivation layer. The input layer corresponds to the state vector dimension, and the output layer corresponds to the action space dimension.

[0024] The first deep Q-network is responsible for action selection, generating a set of candidate actions with the highest Q-values ​​and calculating the Q-value confidence interval for each candidate action. The final action is selected based on an upper confidence bound strategy. The second deep Q-network is responsible for independently evaluating the selected action and calculating the Q-value. The evaluation results of the two networks are used to calculate the Q-value difference. When the Q-value difference exceeds a preset threshold, an uncertainty warning is triggered. The final action evaluation value is obtained by dynamically weighting the Q-values ​​of the two networks. The dynamic weighting coefficients are determined based on the historical prediction accuracy of each network.

[0025] A comprehensive reward function is constructed, comprising rewards for improving physiological indicators, penalties for adverse reactions when multiple drugs are used simultaneously, dosing safety constraints to ensure that the dosage is within a safe range, rewards for temporal rationality to ensure the rationality of the time interval between adjacent dosings, and rewards for smoothness of dosing dosage. The weight coefficients of the comprehensive reward function are determined using a Bayesian optimization method. The Bayesian optimization method includes establishing a Gaussian process regression model to predict the objective function value of the weight coefficient combination, selecting the next set of weight coefficient combinations to be evaluated based on the maximization expectation improvement criterion, and iteratively optimizing until convergence. The objective function comprehensively considers the accuracy of the dosing regimen, the incidence of adverse reactions, and the stability index of the regimen.

[0026] The dual-depth Q-network is trained based on a priority experience replay mechanism, and supervised pre-training is performed using expert knowledge during the training process. An initial dosing regimen is generated based on the trained dual-depth Q-network.

[0027] In one alternative implementation,

[0028] The energy management unit harvests and stores energy using flexible photovoltaic thin films and phase change energy storage materials, and establishes a multi-objective optimization function based on drug preservation conditions and environmental parameters to intelligently allocate and adaptively adjust the power consumption of the infusion pump system. The steps include:

[0029] Energy is harvested by a flexible photovoltaic film on the surface of an intelligent infusion pump in the field. The energy harvested by the flexible photovoltaic film is then input into a microcapsule phase change energy storage material with octadecyl n-alkane as the core for storage. The microcapsule phase change energy storage material has a wall material with controllable cross-linking degree, which is used for phase change energy storage and provides a constant temperature environment for drug preservation.

[0030] Based on the real-time photovoltaic power of the flexible photovoltaic film, the energy storage power of the microcapsule phase change energy storage material, the real-time temperature of the drug, and the target temperature, a multi-objective optimization function is constructed. The multi-objective optimization function includes an energy balance objective function, a temperature control objective function, and a power consumption allocation objective function.

[0031] An improved NSGA-III algorithm with an adaptive crossover operator is used to solve the multi-objective optimization function. The crossover probability of the adaptive crossover operator is related to the fitness value of the current solution, and an adaptive update mechanism for reference points and a local search strategy are introduced.

[0032] Based on the solution results of the multi-objective optimization function, the power consumption of the injection pump system is dynamically allocated in a hierarchical manner according to the priority order of the core functional layer, the auxiliary functional layer and the emergency backup layer. The core functional layer includes a micro peristaltic pump mechanism, a multi-sensor array and a control processor, and the auxiliary functional layer includes a temperature control module and a communication module.

[0033] Based on weather conditions, the system predicts future light intensity and calculates available power. A fuzzy PID controller is constructed by combining light intensity change rate, battery state of charge, and drug temperature deviation. This controller outputs power adjustment values ​​for each functional module, which includes a micro-peristaltic pump mechanism, a multi-sensor array, a control processor, a temperature control module, and a communication module. A weighted average of the current target power and historical power is used to obtain a transitional power value. Power is allocated to each functional module based on this transitional power value. The system monitors the power utilization rate and actual power value of each functional module in real time. Power reallocation is triggered when the difference between the maximum and minimum power utilization rate exceeds a preset adjustment threshold, or when the deviation between the actual power value and the target power value exceeds a preset range.

[0034] In one alternative implementation,

[0035] The improved NSGA-III algorithm with an adaptive crossover operator is used to solve the multi-objective optimization function. The crossover probability of the adaptive crossover operator is related to the fitness value of the current solution. The steps of introducing an adaptive update mechanism for reference points and a local search strategy include:

[0036] Improvements to the NSGA-III algorithm include the introduction of an adaptive crossover operator, an adaptive update mechanism for reference points, and a local search strategy;

[0037] A calculation formula for an adaptive crossover operator is constructed, which includes the baseline crossover probability, the normalized fitness value of the current solution, the maximum fitness value of the population, the current iteration number, and the adaptive adjustment coefficient. The crossover probability is dynamically adjusted according to the calculation formula. The crossover probability is inversely proportional to the fitness value of the current solution and inversely proportional to the iteration number.

[0038] Based on the crossover operation results of the adaptive crossover operator, the reference point is dynamically updated using an exponential moving average method. The position vector of the reference point, the centroid of the non-dominated solution set, and the learning rate are used as parameters to calculate the new reference point position. The distance and average distance between adjacent reference points are calculated, and a distance-based density penalty factor is constructed. When the density penalty factor is greater than the reference point splitting threshold, adjacent reference points are selected in the high-density region to calculate the midpoint position and generate a new reference point. Based on the reference point update results, the initial search radius, the current generation, the maximum generation, and the fitness value are used as parameters to construct a dynamic adjustment formula for the search radius. Based on the gradient information of the current solution, a probability distribution of the search direction is constructed to obtain the local search direction.

[0039] The population diversity index is calculated based on the local search direction. The population diversity index is compared with the local search trigger threshold and the trigger state of the local search is determined by combining the iteration interval. The local search operation is performed according to the trigger state to obtain the local search result. The local search result is combined with the population high-quality solution retention mechanism to design a population re-initialization strategy. The optimized population is obtained by retaining high-quality solutions, random initialization and local search.

[0040] Monitor the stability status of the injection pump system, adaptively adjust the optimization period based on the stability status of the injection pump system, and update the optimized population based on the adjusted optimization period.

[0041] In one alternative implementation,

[0042] The steps for assessing the status of the infusion pump tubing based on multimodal sensor information, performing dynamic safety assessments in conjunction with real-time physiological monitoring, and triggering tiered protection include:

[0043] Real-time data from pressure, bubble, and temperature sensors in the intelligent injection module are collected. This real-time data is input into a dual-stream attention network for feature extraction. Wavelet transform is used to perform multi-scale decomposition of the pressure data, and a target detection network is used to identify and locate bubble data. Temperature data features are obtained through adaptive piecewise fitting, resulting in multimodal sensing features. These multimodal sensing features are then input into a cross-modal self-attention network to calculate the correlation weights between features and perform feature fusion. A variational autoencoder is used to learn sample distribution features, and an anomaly score is calculated using an isolated forest algorithm to generate anomaly warning information for the injection pump tubing status. Real-time physiological parameters are constructed into a graph network, and associated features are extracted through graph convolution operations. A multi-head recursive attention mechanism is introduced to obtain temporal change features. The anomaly warning information is fused with these temporal change features to generate hierarchical safety features.

[0044] Based on the aforementioned hierarchical security features, an adaptive fuzzy inference system is constructed. A particle swarm optimization algorithm is used to dynamically optimize the fuzzy rules, and an adversarial training strategy is introduced. The system performs inference operations on the current state based on the optimized fuzzy rules to obtain a security risk score, which is then divided into multiple security risk levels. Based on these security risk levels, a hierarchical reinforcement learning framework is used for protection decisions. The higher-level network maps different risk levels to corresponding protection measures, while the lower-level network converts these protection measures into execution instructions. An adaptive PID controller is constructed based on these execution instructions, and a model predictive control algorithm is used to adjust the PID parameters in real time, outputting protection control instructions of the corresponding level.

[0045] A second aspect of the present invention,

[0046] Provides an automated control system for intelligent infusion pumps in the field based on physiological parameter monitoring, including:

[0047] The first unit is used to collect patient heart rate, respiration, blood pressure, blood oxygen saturation and body surface electrical signal information through a flexible electronic patch. The flexible electronic patch has a built-in signal processing chip to perform signal conditioning and noise reduction processing on the collected information to obtain physiological parameter information.

[0048] The second unit is used to input physiological parameter information into the intelligent decision-making module. The intelligent decision-making module includes a physiological parameter analysis unit and a medication optimization unit. The physiological parameter analysis unit performs multimodal feature extraction and trend prediction on the physiological parameter information to generate a physiological function status report. The medication optimization unit generates a personalized dosing plan based on a reinforcement learning algorithm combined with the physiological function status report and a drug action mechanism knowledge base. The personalized dosing plan includes drug ratio, injection rate, and injection sequence.

[0049] The third unit is used to send the personalized drug delivery plan to the intelligent injection module. The intelligent injection module includes a micro peristaltic pump mechanism, a multi-sensor array, a control processor, and an energy management unit. The control processor adaptively compensates the injection parameters based on the ambient temperature, humidity, and air pressure parameters collected in real time by the multi-sensor array, and executes the compensated injection parameters through the micro peristaltic pump mechanism. The energy management unit collects and stores energy through flexible photovoltaic thin films and phase change energy storage materials, and establishes a multi-objective optimization function based on drug storage conditions and environmental parameters to intelligently allocate and adaptively adjust the power consumption of the injection pump system. It evaluates the status of the injection pump tubing based on multimodal sensor information, performs dynamic safety assessment in conjunction with real-time physiological monitoring, and triggers graded protection.

[0050] A third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] Memory used to store processor-executable instructions;

[0054] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0055] Fourth aspect of the present invention,

[0056] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0057] This invention uses a flexible electronic patch to collect multiple physiological parameters in real time and utilizes a built-in signal processing chip for signal conditioning and noise reduction. This allows for accurate acquisition of the patient's physiological state information, providing a reliable data foundation for intelligent decision-making and effectively improving the accuracy and reliability of physiological parameter monitoring.

[0058] This invention employs an intelligent decision-making module to extract multimodal features and predict trends of physiological parameters. Combined with reinforcement learning algorithms and a drug action mechanism knowledge base, it generates personalized dosing regimens, thereby achieving intelligent and personalized dosing. The dosing strategy can be dynamically adjusted according to the patient's actual physiological state, improving treatment efficacy and medication safety.

[0059] The intelligent injection module of this invention monitors environmental parameters in real time through a multi-sensor array, adaptively compensates for injection parameters, and employs flexible photovoltaic thin films and phase change energy storage materials for energy management, achieving self-powering and intelligent energy saving. Simultaneously, it assesses the status of the injection pump tubing based on multimodal sensing information, and combines real-time physiological monitoring to achieve dynamic safety assessment and graded protection, significantly improving the reliability and safety of the injection pump system in field environments. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the automatic control method for an intelligent infusion pump in the field based on physiological parameter monitoring, according to an embodiment of the present invention.

[0061] Figure 2 This is a schematic diagram of the automatic control system for an intelligent infusion pump in the field based on physiological parameter monitoring, according to an embodiment of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0064] Figure 1 This is a flowchart illustrating the automatic control method for an intelligent infusion pump in the field based on physiological parameter monitoring, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0065] S1. The patient's heart rate, respiration, blood pressure, blood oxygen saturation and body surface electrical signal information are collected through a flexible electronic patch. The flexible electronic patch has a built-in signal processing chip to perform signal conditioning and noise reduction processing on the collected information to obtain physiological parameter information.

[0066] S2. Input physiological parameter information into the intelligent decision-making module. The intelligent decision-making module includes a physiological parameter analysis unit and a medication optimization unit. The physiological parameter analysis unit performs multimodal feature extraction and trend prediction on the physiological parameter information to generate a physiological function status report. The medication optimization unit generates a personalized dosing plan based on a reinforcement learning algorithm, combined with the physiological function status report and a drug action mechanism knowledge base. The personalized dosing plan includes drug ratio, injection rate, and injection sequence.

[0067] S3. The personalized drug delivery plan is sent to the intelligent injection module, which includes a micro peristaltic pump mechanism, a multi-sensor array, a control processor, and an energy management unit. The control processor adaptively compensates for the injection parameters based on the ambient temperature, humidity, and air pressure parameters collected in real time by the multi-sensor array, and executes the compensated injection parameters through the micro peristaltic pump mechanism. The energy management unit collects and stores energy through flexible photovoltaic thin films and phase change energy storage materials, and establishes a multi-objective optimization function based on drug preservation conditions and environmental parameters to intelligently allocate and adaptively adjust the power consumption of the injection pump system. The unit evaluates the status of the injection pump tubing based on multimodal sensor information, performs dynamic safety assessment in conjunction with real-time physiological monitoring, and triggers graded protection.

[0068] In one alternative implementation,

[0069] The steps for generating a physiological function status report by performing multimodal feature extraction and trend prediction on the physiological parameter information through the physiological parameter analysis unit include:

[0070] Multimodal feature extraction is performed on the physiological parameter information. Temporal morphological features are extracted using a deep convolutional network, and frequency domain features are extracted using a fast Fourier transform and a spectral encoder. Nonlinear dynamic features are extracted by calculating sample entropy and Lyapunov exponent. The deep convolutional network uses batch normalization and ReLU activation function, and the spectral encoder uses layer normalization. A bidirectional attention-enhanced long short-term memory network is used to model the temporal dependency of the multimodal features. The long short-term memory network includes a bidirectional long short-term memory layer for forward and backward information fusion, a multi-head self-attention layer for establishing temporal correlation weights, and a residual connection layer for preserving the original features.

[0071] A multi-scale trend prediction framework is constructed based on the results of temporal dependency modeling, including a short-term predictor based on gated recurrent unit network for five-minute scale prediction, a medium-term predictor based on Transformer encoder for thirty-minute scale prediction, and a long-term predictor based on temporal convolutional network for four-hour scale prediction.

[0072] The multi-scale trend prediction results are input into the cardiovascular system scoring unit and the respiratory system scoring unit based on the gated attention mechanism to score physiological indicators. The cardiovascular system scoring unit calculates the score based on heart rate variability, blood pressure fluctuation characteristics and body surface electrical signal morphology characteristics. The respiratory system scoring unit calculates the score based on respiratory rate changes, blood oxygen saturation trends and respiratory rhythm characteristics.

[0073] Trend analysis was performed on the physiological index scores. The trend analysis used an ensemble learning strategy to fuse the prediction results of the autoregressive moving average model and the Prophet model, and the confidence interval of the prediction results was calculated using the Monte Carlo sampling method.

[0074] A physiological function status report is generated based on the confidence intervals of the physiological indicator scores, trend analysis, and prediction results. The physiological function status report includes cardiovascular system stability analysis results, respiratory system dynamic characteristics, autonomic nervous system regulation capacity indicators, organ function status quantification results, and physiological rhythm homeostasis evaluation.

[0075] For example, the first step is to extract multimodal features from the collected physiological parameter information. This step includes three main aspects: temporal morphological feature extraction, frequency domain feature extraction, and nonlinear dynamic feature extraction.

[0076] For temporal morphological feature extraction, a deep convolutional neural network is used. This network consists of multiple convolutional layers, pooling layers, and fully connected layers. Batch normalization is applied after each convolutional layer to accelerate network convergence and improve model generalization ability. The ReLU activation function is chosen to effectively alleviate the vanishing gradient problem. The network input is raw physiological signals, such as electrocardiograms and blood pressure waveforms. Through multi-layer convolutional operations, the network can automatically learn and extract the temporal morphological features of the signals, such as the morphology of the QRS complex and the amplitude of the T wave.

[0077] For frequency domain feature extraction, the original signal is first subjected to a Fast Fourier Transform (FFT) to convert the time-domain signal into a frequency-domain representation. Then, a spectrum encoder is used to further process the frequency-domain information. The spectrum encoder employs a multilayer perceptron structure, with layer normalization applied after each layer to maintain the relative importance of information in different frequency bands. The encoder output is the compressed frequency-domain feature representation, containing information such as the signal's frequency composition and dominant frequency component.

[0078] Nonlinear dynamics feature extraction is primarily achieved by calculating sample entropy and the Lyapunov exponent. Sample entropy reflects the complexity and irregularity of the signal, while the Lyapunov exponent characterizes the degree of chaos in the system. These two indices can describe the intrinsic properties of physiological systems from the perspective of nonlinear dynamics.

[0079] Next, a bidirectional attention-enhanced long short-term memory network is used to model the temporal dependencies of the extracted multimodal features. This network structure consists of three key components: a bidirectional long short-term memory layer, a multi-head self-attention layer, and a residual connection layer.

[0080] The bidirectional Long Short-Term Memory (LSTM) layer consists of two LSTM units, one forward and one backward, capable of simultaneously considering past and future contextual information. The forward LSTM processes the sequence from left to right, capturing dependencies from the past to the present; the backward LSTM processes the sequence from right to left, capturing dependencies from the future to the present. The outputs from both directions are fused through a concatenation operation to form a feature representation containing bidirectional information.

[0081] Multi-head self-attention layers are used to establish temporal relational weights. They map input features to three spaces: query, key, and value, and then calculate the similarity between the query and key to obtain the attention weights. By using multiple attention heads, temporal dependencies can be learned from different representation subspaces. Finally, the outputs of the multiple heads are concatenated and subjected to a linear transformation to obtain the final attention-enhanced features.

[0082] The role of residual connection layers is to preserve the original feature information. They directly add the input features to the output of the attention layer, helping to alleviate the vanishing gradient problem during deep network training while maintaining the important information of the original features.

[0083] Based on the results of time-series dependency modeling, a multi-scale trend prediction framework is constructed. This framework includes three predictors, each used for prediction tasks at different time scales.

[0084] The short-term predictor is implemented based on a gated recurrent unit (GRU) network for five-minute scale predictions. The GRU network has update and reset gates, effectively capturing rapidly changing trends in the short term. The network input is a feature sequence from the past hour, and the output is a prediction for the next five minutes.

[0085] The intermediate-term predictor employs a Transformer encoder architecture for predictions on a 30-minute timescale. The Transformer's self-attention mechanism can establish long-range dependencies, making it suitable for capturing intermediate-term trends. The encoder stacks multiple layers of self-attention and feedforward networks, taking the feature sequence from the past four hours as input and outputting the predicted values ​​for the next 30 minutes.

[0086] The long-term predictor is designed based on a Temporal Convolutional Network (TCN) for predictions on a four-hour timescale. TCN achieves a large receptive field through causal convolution and dilated convolution, effectively modeling long-term temporal dependencies. The network input is the feature sequence of the past 24 hours, and the output is the predicted value for the next four hours.

[0087] The multi-scale trend prediction results are input into a scoring unit based on a gated attention mechanism to score the cardiovascular and respiratory systems respectively.

[0088] The cardiovascular scoring unit considers three main aspects: heart rate variability, blood pressure variability, and surface electrical signal morphology. Heart rate variability includes time-domain metrics such as SDNN (Standard Deviation of Normal-to-Normal intervals, the total RR intervals of sinus beats) and RMSSD (root mean square of successive differences between adjacent RR intervals), as well as frequency-domain metrics such as the LF / HF ratio (the ratio of low-frequency power to high-frequency power). Blood pressure variability considers the amplitude and frequency of changes in systolic and diastolic blood pressure. Surface electrical signal morphology primarily focuses on ST segment changes and T wave morphology on the electrocardiogram. These features are weighted and fused using a gated attention mechanism to obtain the final cardiovascular score.

[0089] The respiratory system scoring unit is primarily based on changes in respiratory rate, trends in blood oxygen saturation, and respiratory rhythm characteristics. Changes in respiratory rate reflect the respiratory system's regulatory capacity, trends in blood oxygen saturation characterize the efficiency of gas exchange in the lungs, and respiratory rhythm characteristics describe the regularity and depth of respiration. These characteristics are weighted using a gated attention mechanism to calculate a comprehensive respiratory system score.

[0090] Trend analysis was performed on the physiological indicator scores, and an ensemble learning strategy was used to fuse the prediction results of the Autoregressive Moving Average (ARIMA) model and the Prophet model. The ARIMA model is good at capturing linear trends and periodic changes in time series, while the Prophet model is more suitable for handling time series with multiple periodicities and holiday effects. By weighting the prediction results of the two models, a more robust trend prediction can be obtained.

[0091] To quantify the uncertainty of the prediction results, a Monte Carlo sampling method is used to calculate the confidence interval of the prediction results. Specifically, the model parameters are randomly sampled multiple times, and a set of prediction results is generated for each sample. By statistically analyzing the distribution of these prediction results, the confidence interval of the prediction value can be obtained, and the 95% confidence interval is usually selected as the final output.

[0092] Finally, based on physiological index scores, trend analysis, and confidence intervals for prediction results, a comprehensive physiological function status report is generated. This report includes the following key components:

[0093] Cardiovascular system stability analysis results: Comprehensive assessment of heart rate variability, blood pressure regulation capacity and electrocardiogram morphological changes, providing a cardiovascular system stability rating and risk warning.

[0094] Respiratory system dynamic characteristics: Analyze the patterns of change in respiratory rate, depth, and rhythm to assess the adaptability and potential abnormalities of the respiratory system.

[0095] Autonomic nervous system regulation capacity indicators: Based on a comprehensive analysis of heart rate variability and blood pressure variability, the balance between the sympathetic and parasympathetic nervous systems is assessed.

[0096] Organ function status quantification results: Through integrated analysis of multimodal physiological parameters, the functional status of important organs such as the heart, lungs, liver, and kidneys is quantitatively assessed.

[0097] Physiological rhythm homeostasis assessment: Analyze the diurnal rhythm changes of various physiological indicators to assess whether the body's physiological rhythms are in a stable state.

[0098] This invention comprehensively captures the temporal, frequency, and nonlinear dynamic characteristics of physiological signals through multimodal feature extraction technology, providing a rich information foundation for subsequent analysis. The use of a combination of deep learning and attention mechanisms for time-dependent modeling effectively captures the complex interactions and short- and long-term dependencies between physiological parameters, improving the model's expressive power and predictive accuracy. The constructed multi-scale trend prediction framework can simultaneously predict short-, medium-, and long-term physiological state changes, providing comprehensive trend information and facilitating the timely identification of potential risks and the development of intervention strategies. The gated attention-based scoring unit can adaptively adjust the importance of different features, improving the accuracy and interpretability of the scoring. Employing ensemble learning strategies and Monte Carlo sampling methods for trend analysis and uncertainty quantification not only provides reliable prediction results but also gives confidence intervals, offering more comprehensive reference information for clinical decision-making. The generated physiological function status report covers the assessment results of multiple important physiological systems, providing a comprehensive and intuitive overview of the patient's condition.

[0099] In one alternative implementation,

[0100] The medication optimization unit generates a personalized dosing regimen based on a reinforcement learning algorithm combined with the physiological function status report and the drug action mechanism knowledge base. The personalized dosing regimen includes steps such as drug ratio, injection rate, and injection sequence:

[0101] A drug action mechanism knowledge base is constructed. Drug molecular structures are encoded using an attention-enhanced graph neural network to obtain structured representation vectors. Transduction learning is then used to transfer knowledge of known drug action mechanisms to new drugs, forming an entity relationship network containing information on drug molecular structures, targets, metabolic pathways, and drug interactions. Indicators in physiological function status reports are standardized and dimensionality-reduced preprocessing is performed. Historical dosing records and physiological indicator predictions are combined to form a dosing environment state vector. An action space is designed, comprising discretizing drug combinations into standard ratio schemes to form a drug ratio space, discretizing injection rate ranges into multiple levels to form an injection rate space, and dividing dosing time into time segments to form an injection time sequence space.

[0102] By combining the drug action mechanism knowledge base and the drug administration environment state vector, two identical but parameter-independent dual deep Q-networks are constructed. One network is used to select the dosing regimen, and the other network is used to evaluate the Q value of the selected dosing regimen. A comprehensive reward function is designed, which includes physiological indicator improvement rewards, drug interaction penalties, and dosing safety constraints. The weight coefficients of the comprehensive reward function are determined by grid search optimization. An initial dosing regimen is generated based on the trained deep Q-network. A regimen safety assurance mechanism is constructed. The contribution of indicators in the physiological function status report, historical dosing records, and predicted changes in physiological indicators to the Q value is analyzed using SHAP values. The safety of the initial dosing regimen is verified based on the Monte Carlo method. The risk of the initial dosing regimen is assessed by sampling the state transition probability. When a potential risk is detected, an early warning mechanism is activated and the parameters of the initial dosing regimen are automatically adjusted.

[0103] Collect physiological response data to update the drug administration environment state vector, calculate the actual comprehensive reward function value and store it in the experience replay pool, sample data from the experience replay pool to update the strategy network, and when a change in physiological state is detected to exceed a preset change threshold, trigger the drug administration plan adjustment and re-execute the plan safety assurance mechanism to generate the final personalized drug administration plan.

[0104] For example, a drug action mechanism knowledge base is first constructed. This knowledge base processes drug molecular structure information through an attention-enhanced graph neural network, transforming the molecular structure into structured representation vectors. Specifically, a multi-head attention mechanism is used to extract features from atoms and chemical bonds in the molecule, with 8 attention heads and a hidden layer dimension of 256. Through transduction learning, the system can transfer knowledge of known drug action mechanisms to new drugs, achieving a transfer rate of 0.85. The resulting entity relationship network includes information on drug molecular structure, target sites, metabolic pathways, and drug interactions.

[0105] For processing physiological function status reports, the system employs standardization and dimensionality reduction preprocessing methods. During standardization, all physiological indicators are uniformly transformed to a zero-mean, unit variance distribution. Dimensionality reduction utilizes principal component analysis, retaining principal components that explain at least 95% of the variance. Combining historical dosing records, the system uses a 24-hour sliding window to extract features, predicting the physiological indicator trends for the next 4 hours, and forming a 128-dimensional dosing environment state vector.

[0106] The action space design includes three dimensions: the drug ratio space discretizes commonly used drug combinations into 50 standard ratio schemes; the injection rate space discretizes the rate range into 10 levels, with a minimum rate of 0.1 ml / min and a maximum rate of 10 ml / min; and the injection time sequence space divides 24 hours into 144 10-minute time segments.

[0107] The dual deep Q-networks are constructed using the same network structure, consisting of four fully connected layers with 512, 256, 128, and 64 neurons per layer, respectively. The two networks are trained independently, with the policy network selecting the dosing regimen and the target network evaluating the Q-value of the regimen. The comprehensive reward function comprises three components: a reward weight of 0.5 for improvement in physiological indicators, a penalty weight of 0.3 for drug interactions, and a dosing safety constraint weight of 0.2.

[0108] In the safety assurance mechanism of the protocol, the SHAP (SHapley Additive exPlanations) value analysis uses 1000 samples to calculate feature importance. The Monte Carlo method performs 10,000 sample simulations to assess the risk probability of the dosing regimen. When the risk probability exceeds a set threshold of 0.05, an early warning mechanism is triggered, and the system automatically adjusts the dosing parameters, including reducing the injection rate or adjusting the dosing sequence.

[0109] The system collects physiological response data every 5 minutes to update the state vector and stores the actual reward value in an experience replay pool with a capacity of 100,000. Each update randomly samples 256 samples from the experience pool to train the policy network. When a change in physiological indicators exceeds a preset threshold, the system triggers a protocol adjustment process, re-executes the safety assurance mechanism, and generates an updated personalized dosing protocol.

[0110] This invention constructs a complete knowledge base of drug action mechanisms, combines attention-enhanced graph neural networks and transduction learning methods to achieve accurate prediction of the action mechanisms of new drugs, thereby improving the scientific rigor and reliability of dosing regimens. Employing a dual-depth Q-network structure and a comprehensive reward function design, the system can balance multiple dimensions of therapeutic effect, drug interaction, and safety, ensuring the overall optimality of the dosing regimen and reducing the risk of adverse reactions. It innovatively introduces a regimen safety assurance mechanism, using SHAP value analysis and Monte Carlo methods to perform multiple verifications of the dosing regimen, and can rapidly adjust regimen parameters based on real-time physiological responses, ensuring dynamic optimization and safety during the dosing process.

[0111] In one alternative implementation,

[0112] By combining the drug action mechanism knowledge base and the drug administration environment state vector, two identical but parameter-independent dual deep Q-networks are constructed. One network is used to select the dosing regimen, and the other network is used to evaluate the Q-value of the selected regimen. A comprehensive reward function is designed, including rewards for improving physiological indicators, penalties for drug interactions, and dosing safety constraints. The weight coefficients of the comprehensive reward function are determined through grid search optimization. The steps for generating an initial dosing regimen based on the trained deep Q-network include:

[0113] Two identical but parameter-independent dual-depth Q-networks are constructed. Each dual-depth Q-network includes an input layer, three fully connected layers, a batch normalization layer, and a random deactivation layer. The input layer corresponds to the state vector dimension, and the output layer corresponds to the action space dimension.

[0114] The first deep Q-network is responsible for action selection, generating a set of candidate actions with the highest Q-values ​​and calculating the Q-value confidence interval for each candidate action. The final action is selected based on an upper confidence bound strategy. The second deep Q-network is responsible for independently evaluating the selected action and calculating the Q-value. The evaluation results of the two networks are used to calculate the Q-value difference. When the Q-value difference exceeds a preset threshold, an uncertainty warning is triggered. The final action evaluation value is obtained by dynamically weighting the Q-values ​​of the two networks. The dynamic weighting coefficients are determined based on the historical prediction accuracy of each network.

[0115] A comprehensive reward function is constructed, comprising rewards for improving physiological indicators, penalties for adverse reactions when multiple drugs are used simultaneously, dosing safety constraints to ensure that the dosage is within a safe range, rewards for temporal rationality to ensure the rationality of the time interval between adjacent dosings, and rewards for smoothness of dosing dosage. The weight coefficients of the comprehensive reward function are determined using a Bayesian optimization method. The Bayesian optimization method includes establishing a Gaussian process regression model to predict the objective function value of the weight coefficient combination, selecting the next set of weight coefficient combinations to be evaluated based on the maximization expectation improvement criterion, and iteratively optimizing until convergence. The objective function comprehensively considers the accuracy of the dosing regimen, the incidence of adverse reactions, and the stability index of the regimen.

[0116] The dual-depth Q-network is trained based on a priority experience replay mechanism, and supervised pre-training is performed using expert knowledge during the training process. An initial dosing regimen is generated based on the trained dual-depth Q-network.

[0117] For example, when constructing a dual-depth Q-network system based on a drug action mechanism knowledge base and a drug administration environment state vector, the first step is to construct the input data structure. The state vector contains information such as patient physiological indicators, current medication status, and environmental parameters. This data is standardized and then input into the network.

[0118] The dual-depth Q-network employs a symmetric architecture. The number of nodes in the input layer matches the dimension of the state vector. The first hidden layer uses 512 neurons, the second uses 256 neurons, and the third uses 128 neurons. Each hidden layer is followed by a batch normalization layer with a momentum coefficient of 0.99 and an ε value of 0.001. The dropout rate of the random deactivation layer is set to 0.3 to prevent overfitting. The output layer dimension corresponds to the number of selectable dosing regimens.

[0119] The action selection network employs a confidence upper bound method when generating candidate actions. Specifically, it obtains the Q-value distribution for each action through Monte Carlo sampling, calculates the mean and standard deviation, and selects the top K actions with the highest confidence upper bound as the candidate set. The value of K can be set according to actual needs, typically 5-10. The evaluation network independently calculates the Q-values ​​of these candidate actions, issuing a warning when the difference between the Q-values ​​of the two networks exceeds a set threshold (e.g., 20%). The final action evaluation value is obtained by a weighted average of the Q-values ​​of the two networks, with the weight coefficients dynamically adjusted based on the mean squared error of each network's most recent 100 predictions.

[0120] The construction of the comprehensive reward function considers multiple aspects. The reward for improvement in physiological indicators is calculated based on the magnitude of change in key indicators; the penalty for drug interactions is obtained by querying a drug interaction database to obtain risk coefficients; dosing safety constraints ensure that the dosage is within the recommended range; the reward for reasonable timing ensures that the interval between adjacent dosing doses is not less than the minimum safe interval; and the reward for dose smoothness controls the dose change gradient. When using Bayesian optimization to determine the weight coefficients, the kernel function of the Gaussian process regression model is set to a radial basis function, the length scale parameter is set to 1.0, and the signal variance is set to 1.0.

[0121] During training, the priority experience replay buffer size was set to 100,000, and the batch size for each sampling was 64. The priority of new samples was calculated based on a time decay factor of 0.99 and TD error. Clinically recommended dosing regimens were used in the expert knowledge pre-training phase, with 50 pre-training rounds.

[0122] This invention achieves reliability and stability in dosing regimen selection by constructing a dual-depth Q-network structure, reducing the decision-making risks that a single network might bring, and improving the accuracy and credibility of the dosing regimen. Employing a comprehensive reward function and Bayesian optimization method, it fully considers multiple dimensions such as physiological indicator improvement, drug interactions, and dosing safety, ensuring that the generated dosing regimen not only guarantees therapeutic efficacy but also minimizes the risk of adverse reactions. Combining a priority experience replay mechanism and an expert knowledge pre-training strategy accelerates model convergence, improves training efficiency, and ensures consistency between the generated regimen and clinical practice, enhancing the model's practicality and generalizability.

[0123] In one alternative implementation,

[0124] The energy management unit harvests and stores energy using flexible photovoltaic thin films and phase change energy storage materials, and establishes a multi-objective optimization function based on drug preservation conditions and environmental parameters to intelligently allocate and adaptively adjust the power consumption of the infusion pump system. The steps include:

[0125] Energy is harvested by a flexible photovoltaic film on the surface of an intelligent infusion pump in the field. The energy harvested by the flexible photovoltaic film is then input into a microcapsule phase change energy storage material with octadecyl n-alkane as the core for storage. The microcapsule phase change energy storage material has a wall material with controllable cross-linking degree, which is used for phase change energy storage and provides a constant temperature environment for drug preservation.

[0126] Based on the real-time photovoltaic power of the flexible photovoltaic film, the energy storage power of the microcapsule phase change energy storage material, the real-time temperature of the drug, and the target temperature, a multi-objective optimization function is constructed. The multi-objective optimization function includes an energy balance objective function, a temperature control objective function, and a power consumption allocation objective function.

[0127] An improved NSGA-III algorithm with an adaptive crossover operator is used to solve the multi-objective optimization function. The crossover probability of the adaptive crossover operator is related to the fitness value of the current solution, and an adaptive update mechanism for reference points and a local search strategy are introduced.

[0128] Based on the solution results of the multi-objective optimization function, the power consumption of the injection pump system is dynamically allocated in a hierarchical manner according to the priority order of the core functional layer, the auxiliary functional layer and the emergency backup layer. The core functional layer includes a micro peristaltic pump mechanism, a multi-sensor array and a control processor, and the auxiliary functional layer includes a temperature control module and a communication module.

[0129] Based on weather conditions, the system predicts future light intensity and calculates available power. A fuzzy PID controller is constructed by combining light intensity change rate, battery state of charge, and drug temperature deviation. This controller outputs power adjustment values ​​for each functional module, which includes a micro-peristaltic pump mechanism, a multi-sensor array, a control processor, a temperature control module, and a communication module. A weighted average of the current target power and historical power is used to obtain a transitional power value. Power is allocated to each functional module based on this transitional power value. The system monitors the power utilization rate and actual power value of each functional module in real time. Power reallocation is triggered when the difference between the maximum and minimum power utilization rate exceeds a preset adjustment threshold, or when the deviation between the actual power value and the target power value exceeds a preset range.

[0130] For example, the energy management unit of the field intelligent injection pump first harvests energy through a flexible photovoltaic film applied to the outer surface of the pump. This flexible photovoltaic film, made of a polymer substrate, is 0.5 mm thick and exhibits good flexibility and photoelectric conversion efficiency. Under standard illumination conditions, each square centimeter of photovoltaic film can generate 2.5 milliwatts of output power. The harvested energy is then stored in a microcapsule phase change energy storage material. The microcapsules use octadecyl n-alkane as the core material, with a phase change temperature of 28 degrees Celsius and a latent heat value of 240 joules / gram. The microcapsule wall material is a modified polyacrylate material, and the degree of crosslinking can be controlled between 15% and 45% by adjusting the amount of crosslinking agent, thus achieving structural stability during the phase change process.

[0131] The system monitors parameters such as photovoltaic power harvesting, energy storage power, and drug temperature in real time, and constructs a multi-objective optimization model. When the light intensity is 800 watts / square meter, the photovoltaic power harvesting can reach 150 milliwatts, and the energy storage material charging and discharging power is 80 milliwatts. The target drug temperature is set at 25 degrees Celsius, with an allowable fluctuation range of ±0.5 degrees Celsius. The optimization model comprehensively considers three objectives: energy balance, temperature control, and power consumption allocation. It is solved by an improved NSGA-III algorithm that introduces an adaptive crossover operator, an adaptive update mechanism for reference points, and a local search strategy. In the algorithm, the adaptive crossover probability is dynamically adjusted between 0.6 and 0.9, the number of reference points is set to 300, and the reference point distribution is updated every 50 generations.

[0132] The system allocates power consumption across three levels for the infusion pump functional module. The core functional layer includes a micro peristaltic pump mechanism (30 mW), a sensor array (15 mW), and a control processor (25 mW). The auxiliary functional layer includes a temperature control module (20 mW) and a communication module (10 mW). An emergency backup layer reserves 20 mW of power. Based on meteorological data predicting changes in light intensity over the next 4 hours, combined with the current battery state of charge (set threshold of 30%) and drug temperature deviation, a fuzzy PID controller dynamically adjusts the power of each module.

[0133] The system samples power data from each module every 5 minutes and calculates transitional power values ​​using the three most recent data points with weights of 0.5, 0.3, and 0.2. A power reallocation process is triggered when the difference between the maximum and minimum power utilization exceeds 20%, or when the actual power deviates from the target power by more than 15%. During reallocation, priority is given to powering the core functional layer, and the operating mode of the auxiliary functional layer is adjusted based on the remaining energy status.

[0134] This invention achieves continuous and stable power supply in field environments through the synergistic effect of flexible photovoltaic thin films and phase change energy storage materials. The phase change energy storage materials simultaneously provide a constant temperature environment for the drugs, improving the system's energy utilization efficiency and drug storage safety. By employing an improved multi-objective optimization algorithm and a hierarchical power allocation strategy, the system achieves dynamic balance of functions under limited energy conditions, ensuring the continuous and stable operation of core functions and improving the system's reliability and adaptability. Based on a predictive power regulation mechanism and adaptive control strategy, the system can adjust the power allocation scheme in a timely manner according to environmental changes and operating status, avoiding the risk of energy depletion, extending field working time, and enhancing the practicality and intelligence level of the infusion pump.

[0135] In one alternative implementation,

[0136] The improved NSGA-III algorithm with an adaptive crossover operator is used to solve the multi-objective optimization function. The crossover probability of the adaptive crossover operator is related to the fitness value of the current solution. The steps of introducing an adaptive update mechanism for reference points and a local search strategy include:

[0137] Improvements to the NSGA-III algorithm include the introduction of an adaptive crossover operator, an adaptive update mechanism for reference points, and a local search strategy;

[0138] A calculation formula for an adaptive crossover operator is constructed, which includes the baseline crossover probability, the normalized fitness value of the current solution, the maximum fitness value of the population, the current iteration number, and the adaptive adjustment coefficient. The crossover probability is dynamically adjusted according to the calculation formula. The crossover probability is inversely proportional to the fitness value of the current solution and inversely proportional to the iteration number.

[0139] Based on the crossover operation results of the adaptive crossover operator, the reference point is dynamically updated using an exponential moving average method. The position vector of the reference point, the centroid of the non-dominated solution set, and the learning rate are used as parameters to calculate the new reference point position. The distance and average distance between adjacent reference points are calculated, and a distance-based density penalty factor is constructed. When the density penalty factor is greater than the reference point splitting threshold, adjacent reference points are selected in the high-density region to calculate the midpoint position and generate a new reference point. Based on the reference point update results, the initial search radius, the current generation, the maximum generation, and the fitness value are used as parameters to construct a dynamic adjustment formula for the search radius. Based on the gradient information of the current solution, a probability distribution of the search direction is constructed to obtain the local search direction.

[0140] The population diversity index is calculated based on the local search direction. The population diversity index is compared with the local search trigger threshold and the trigger state of the local search is determined by combining the iteration interval. The local search operation is performed according to the trigger state to obtain the local search result. The local search result is combined with the population high-quality solution retention mechanism to design a population re-initialization strategy. The optimized population is obtained by retaining high-quality solutions, random initialization and local search.

[0141] Monitor the stability status of the injection pump system, adaptively adjust the optimization period based on the stability status of the injection pump system, and update the optimized population based on the adjusted optimization period.

[0142] For example, firstly, the NSGA-III algorithm is improved. The main improvements include the introduction of an adaptive crossover operator, an adaptive reference point update mechanism, and a local search strategy. These improvements aim to enhance the algorithm's performance and adaptability.

[0143] Secondly, a formula for calculating the adaptive crossover operator is constructed. This formula considers factors such as the baseline crossover probability, the normalized fitness value of the current solution, the maximum fitness value of the population, the current iteration number, and the adaptive adjustment coefficient. According to this formula, the crossover probability dynamically decreases as the fitness value of the current solution and the iteration number increase. For example, assuming the baseline crossover probability is 0.8, the normalized fitness value of the current solution is 0.6, the maximum fitness value of the population is 1.0, the current iteration number is 50, the maximum iteration number is 100, and the adaptive adjustment coefficient is 0.5, the calculated crossover probability might be 0.65.

[0144] Next, based on the crossover operation results of the adaptive crossover operator, the reference point is dynamically updated using an exponential moving average method. This process involves the position vector of the reference point, the centroid of the non-dominated solution set, and the learning rate. For example, if the current reference point position is (0.3, 0.4, 0.3), the centroid of the non-dominated solution set is (0.35, 0.35, 0.3), and the learning rate is 0.1, then the updated reference point position might be (0.305, 0.395, 0.3).

[0145] Then, the distance and average distance between adjacent reference points are calculated to construct a distance-based density penalty factor. When the density penalty factor is greater than a preset reference point splitting threshold, the midpoint position is calculated for adjacent reference points in the high-density region, generating a new reference point. For example, if the average distance between adjacent reference points is 0.1, and the distance between two adjacent reference points is 0.05, the density penalty factor might be 2.0. Assuming the reference point splitting threshold is 1.5, a new reference point will be generated between these two reference points.

[0146] Based on the reference point update results, a dynamic adjustment formula for the search radius is constructed. This formula considers factors such as the initial search radius, the current generation, the maximum generation, and the fitness value. Simultaneously, a probability distribution of the search direction is constructed based on the gradient information of the current solution to obtain the local search direction. For example, if the initial search radius is 0.1, the current generation is 30, the maximum generation is 100, and the fitness value of the current solution is 0.8, then the adjusted search radius might be 0.07.

[0147] Next, a population diversity index is calculated based on the local search direction. This index is compared with a preset local search trigger threshold, and the trigger state of the local search is determined in conjunction with the iteration interval. A local search operation is then performed based on the trigger state to obtain the local search result. For example, if the calculated population diversity index is 0.6, the local search trigger threshold is 0.5, and the current iteration count meets the preset iteration interval, then a local search operation is triggered.

[0148] A population reinitialization strategy is designed by combining local search results with a mechanism for preserving high-quality solutions. The optimized population is obtained through preserving high-quality solutions, random initialization, and local search. For example, the top 20% of high-quality solutions can be preserved, 60% of individuals can be randomly initialized, and the remaining 20% ​​can be obtained through local search.

[0149] Finally, the stability of the infusion pump system is monitored, and the optimization period is adaptively adjusted based on this status. The optimized population is then updated based on the adjusted optimization period. For example, if the system stability is detected to be good, the optimization period can be appropriately extended; conversely, the optimization period can be shortened to respond more quickly to system changes. Through the above steps, the method of the present invention can effectively solve multi-objective optimization problems, and is particularly suitable for the optimization control of infusion pump systems.

[0150] This invention introduces an adaptive crossover operator to dynamically adjust the crossover probability based on the fitness value of the current solution and the iteration process, effectively balancing the algorithm's global search capability and local refinement capability, thus improving the algorithm's convergence speed and solution quality. By employing a reference point adaptive update mechanism and a distance-based density penalty factor, the algorithm can better adapt to complex multi-objective optimization problems, improving its search efficiency in high-dimensional objective spaces while maintaining population diversity. Furthermore, by introducing a local search strategy and a population re-initialization mechanism, combined with monitoring the stability state of the injection pump system and adaptively adjusting the optimization cycle, the invention can improve the accuracy of local searches while ensuring global search capability, and can quickly respond to changes in system state, achieving efficient optimization control of the injection pump system.

[0151] In one alternative implementation,

[0152] The steps for assessing the status of the infusion pump tubing based on multimodal sensor information, performing dynamic safety assessments in conjunction with real-time physiological monitoring, and triggering tiered protection include:

[0153] Real-time data from pressure, bubble, and temperature sensors in the intelligent injection module are collected. This real-time data is input into a dual-stream attention network for feature extraction. Wavelet transform is used to perform multi-scale decomposition of the pressure data, and a target detection network is used to identify and locate bubble data. Temperature data features are obtained through adaptive piecewise fitting, resulting in multimodal sensing features. These multimodal sensing features are then input into a cross-modal self-attention network to calculate the correlation weights between features and perform feature fusion. A variational autoencoder is used to learn sample distribution features, and an anomaly score is calculated using an isolated forest algorithm to generate anomaly warning information for the injection pump tubing status. Real-time physiological parameters are constructed into a graph network, and associated features are extracted through graph convolution operations. A multi-head recursive attention mechanism is introduced to obtain temporal change features. The anomaly warning information is fused with these temporal change features to generate hierarchical safety features.

[0154] Based on the aforementioned hierarchical security features, an adaptive fuzzy inference system is constructed. A particle swarm optimization algorithm is used to dynamically optimize the fuzzy rules, and an adversarial training strategy is introduced. The system performs inference operations on the current state based on the optimized fuzzy rules to obtain a security risk score, which is then divided into multiple security risk levels. Based on these security risk levels, a hierarchical reinforcement learning framework is used for protection decisions. The higher-level network maps different risk levels to corresponding protection measures, while the lower-level network converts these protection measures into execution instructions. An adaptive PID controller is constructed based on these execution instructions, and a model predictive control algorithm is used to adjust the PID parameters in real time, outputting protection control instructions of the corresponding level.

[0155] For example, firstly, real-time data from the pressure sensor, bubble sensor, and temperature sensor in the intelligent injection module are collected. The pressure sensor is a high-precision piezoelectric sensor with a sampling frequency of 1000Hz; the bubble sensor is an ultrasonic sensor with a sampling frequency of 500Hz; and the temperature sensor is a thermocouple with a sampling frequency of 10Hz.

[0156] Next, the collected real-time data is input into a dual-stream attention network for feature extraction. The dual-stream attention network consists of two branches: a spatial stream and a temporal stream. The spatial stream uses a 3D convolutional neural network to extract spatial features, while the temporal stream uses a long short-term memory network to extract temporal features. The outputs of the two branches are fused through an attention mechanism to obtain a preliminary feature representation.

[0157] For the pressure data, wavelet transform was used for multi-scale decomposition. Specifically, the db4 wavelet basis function was used to perform a 5-level decomposition of the pressure data to obtain wavelet coefficients in different frequency bands. By analyzing the energy distribution and statistical characteristics of the wavelet coefficients, the multi-scale features of the pressure data were extracted.

[0158] For bubble data, an object detection network is used for identification and localization. An improved YOLOv5 network is used as the object detection model to perform real-time bubble detection on the images. The network input is a 224x224x3 RGB image, and the output is the bubble's position coordinates and confidence score. The final bubble detection results are filtered using a non-maximum suppression algorithm.

[0159] For temperature data, features are obtained through adaptive piecewise fitting. First, the temperature data is segmented using a sliding window method, with a window size of 60 seconds. Then, a polynomial fitting is applied to each segment, with the fitting order adaptively adjusted from order 1 to 5. By analyzing the fitting parameters and fitting error, the variation characteristics of the temperature data are extracted.

[0160] The aforementioned multimodal sensing features are input into a cross-modal self-attention network to calculate the correlation weights between features and perform feature fusion. The cross-modal self-attention network employs a multi-head attention mechanism with 8 heads. By calculating the similarity between the query vector, key vector, and value vector, the attention weights between features of different modalities are obtained, achieving adaptive feature fusion.

[0161] Next, a variational autoencoder (VAE) is used to learn the sample distribution features. Both the encoder and decoder of the VAE employ a 3-layer fully connected neural network with a hidden layer dimension of 128. The latent space representation of the data is learned by minimizing the reconstruction error and KL divergence. Simultaneously, an anomaly score is calculated using the isolated forest algorithm. The number of trees in the isolated forest is set to 100, and the subsample size is 256. Based on the latent space representation and the anomaly score, anomaly warning information for the status of the injection pump tubing is generated.

[0162] Real-time physiological parameters were constructed as a graph network. These parameters included heart rate, blood pressure, and blood oxygen saturation, with each parameter serving as a node in the graph. The connection weights between nodes were determined based on the Pearson correlation coefficient between the parameters. Associative features were extracted using graph convolution operations, with two convolutional layers and a hidden layer dimension of 64.

[0163] A multi-head recurrent attention mechanism is introduced to capture temporal variation features. A bidirectional gated recurrent unit (GRU) is used as the basic unit of the recurrent neural network, with a hidden layer dimension of 128. The number of heads in the multi-head attention mechanism is set to 4, and long-term dependencies are captured by calculating the attention weights at different time steps.

[0164] Anomaly warning information is fused with temporal variation characteristics to generate hierarchical security features. The fusion method uses weighted summation, with weights automatically learned through backpropagation. The hierarchical security features include security information at three levels: device level, physiological level, and system level.

[0165] An adaptive fuzzy inference system is constructed based on hierarchical security features. The fuzzy inference system contains 5 input variables and 1 output variable, with each variable divided into 3 fuzzy sets. The initial fuzzy rule base contains 243 rules, and the Mamdani inference method is employed.

[0166] A particle swarm optimization (PSO) algorithm is used to dynamically optimize fuzzy rules. The particle swarm size is set to 50, and the maximum number of iterations is 100. Each particle represents a set of parameters for a fuzzy rule, and the rules are optimized by minimizing inference error. Simultaneously, an adversarial training strategy is introduced to enhance the model's robustness by generating adversarial examples.

[0167] The current state is inferred and calculated using optimized fuzzy rules to obtain a safety risk score. The score ranges from 0 to 100, where 0 represents the safest and 100 represents the most dangerous. The safety risk score is divided into three levels: low risk (0-30), medium risk (31-70), and high risk (71-100).

[0168] Based on security risk levels, a hierarchical reinforcement learning framework is used for protection decisions. The higher-level network employs a deep Q-network (DQN), with the state space representing the security risk level and the action space representing different levels of protection measures. The lower-level network uses a policy gradient algorithm to translate protection measures into specific execution instructions.

[0169] Finally, an adaptive PID controller is constructed based on the executed instructions. The initial parameters of the PID controller are set. A model predictive control algorithm is used to adjust the PID parameters in real time, with a prediction time of 10 seconds and a control time of 5 seconds. Optimal PID parameters are obtained by minimizing the prediction error and control variable changes. Finally, the corresponding level of protection control instructions are output to achieve the safety protection of the injection pump.

[0170] This invention achieves comprehensive perception and accurate assessment of the status of infusion pump tubing through multimodal sensor information fusion and deep learning technology. Collaborative analysis of multimodal data improves the accuracy and reliability of anomaly detection, effectively reducing false alarms and false negatives. Dynamic safety assessment, combined with real-time physiological monitoring data, considers the interaction between device status and patient physiological condition, making the safety assessment more comprehensive and personalized. The construction of hierarchical safety features helps analyze potential risks from multiple dimensions, improving the depth and breadth of the safety assessment. Employing a graded protection strategy and adaptive control method, it achieves precise response to different risk levels. The hierarchical reinforcement learning framework can flexibly adjust protective measures according to actual conditions, improving the system's adaptability and scalability. The introduction of an adaptive PID controller ensures the smooth execution of protective commands, effectively enhancing the overall safety performance of the infusion pump system.

[0171] Figure 2 This is a schematic diagram of the automatic control system for an intelligent infusion pump in the field based on physiological parameter monitoring, as described in an embodiment of the present invention. Figure 2 As shown, the system includes:

[0172] The first unit is used to collect patient heart rate, respiration, blood pressure, blood oxygen saturation and body surface electrical signal information through a flexible electronic patch. The flexible electronic patch has a built-in signal processing chip to perform signal conditioning and noise reduction processing on the collected information to obtain physiological parameter information.

[0173] The second unit is used to input physiological parameter information into the intelligent decision-making module. The intelligent decision-making module includes a physiological parameter analysis unit and a medication optimization unit. The physiological parameter analysis unit performs multimodal feature extraction and trend prediction on the physiological parameter information to generate a physiological function status report. The medication optimization unit generates a personalized dosing plan based on a reinforcement learning algorithm combined with the physiological function status report and a drug action mechanism knowledge base. The personalized dosing plan includes drug ratio, injection rate, and injection sequence.

[0174] The third unit is used to send the personalized drug delivery plan to the intelligent injection module. The intelligent injection module includes a micro peristaltic pump mechanism, a multi-sensor array, a control processor, and an energy management unit. The control processor adaptively compensates the injection parameters based on the ambient temperature, humidity, and air pressure parameters collected in real time by the multi-sensor array, and executes the compensated injection parameters through the micro peristaltic pump mechanism. The energy management unit collects and stores energy through flexible photovoltaic thin films and phase change energy storage materials, and establishes a multi-objective optimization function based on drug storage conditions and environmental parameters to intelligently allocate and adaptively adjust the power consumption of the injection pump system. It evaluates the status of the injection pump tubing based on multimodal sensor information, performs dynamic safety assessment in conjunction with real-time physiological monitoring, and triggers graded protection.

[0175] A third aspect of the embodiments of the present invention,

[0176] An electronic device is provided, comprising:

[0177] processor;

[0178] Memory used to store processor-executable instructions;

[0179] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0180] Fourth aspect of the present invention,

[0181] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0182] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic control system for an intelligent infusion pump in the field based on physiological parameter monitoring, characterized in that, include: The first unit is used to collect patient heart rate, respiration, blood pressure, blood oxygen saturation and body surface electrical signal information through a flexible electronic patch. The flexible electronic patch has a built-in signal processing chip to perform signal conditioning and noise reduction processing on the collected information to obtain physiological parameter information. The second unit is used to input physiological parameter information into the intelligent decision-making module. The intelligent decision-making module includes a physiological parameter analysis unit and a medication optimization unit. The physiological parameter analysis unit performs multimodal feature extraction and trend prediction on the physiological parameter information to generate a physiological function status report. The medication optimization unit generates a personalized dosing regimen based on a reinforcement learning algorithm combined with the physiological function status report and a drug action mechanism knowledge base. The personalized dosing regimen includes drug ratio, injection rate, and injection sequence. This includes: constructing a drug action mechanism knowledge base; encoding drug molecular structures using an attention-enhanced graph neural network to obtain drug structured representation vectors; transferring known drug action mechanism knowledge to new drugs using a transduction learning method to form an entity relationship network containing drug molecular structure, target sites, metabolic pathways, and drug interaction information; and designing an action space, which includes discretizing drug combinations into standard ratio schemes to form a drug ratio space, and... The injection rate range is discretized into multiple levels to form an injection rate space, and the dosing time is divided into time segments to form an injection time sequence space. Combining a drug action mechanism knowledge base and a dosing environment state vector, two identical but parameter-independent dual deep Q-networks are constructed. One network is used to select the dosing regimen, and the other network is used to evaluate the Q value of the selected dosing regimen. A comprehensive reward function is designed, including physiological indicator improvement rewards, drug interaction penalties, and dosing safety constraints. The weight coefficients of the comprehensive reward function are determined through grid search optimization. An initial dosing regimen is generated based on the trained deep Q-network. A regimen safety assurance mechanism is constructed. The contribution of indicators in the physiological function status report, historical dosing records, and predicted changes in physiological indicators to the Q value is analyzed using SHAP values. The safety of the initial dosing regimen is verified based on the Monte Carlo method. The risk of the initial dosing regimen is assessed by sampling the state transition probability. When a potential risk is detected, an early warning mechanism is activated and the parameters of the initial dosing regimen are automatically adjusted. The third unit is used to send the personalized drug delivery plan to the intelligent injection module. The intelligent injection module includes a micro peristaltic pump mechanism, a multi-sensor array, a control processor, and an energy management unit. The control processor adaptively compensates the injection parameters based on the ambient temperature, humidity, and air pressure parameters collected in real time by the multi-sensor array, and executes the compensated injection parameters through the micro peristaltic pump mechanism. The energy management unit collects and stores energy through flexible photovoltaic thin films and phase change energy storage materials, and establishes a multi-objective optimization function based on drug storage conditions and environmental parameters to intelligently allocate and adaptively adjust the power consumption of the injection pump system. It evaluates the status of the injection pump tubing based on multimodal sensor information, performs dynamic safety assessment in conjunction with real-time physiological monitoring, and triggers graded protection.

2. The system according to claim 1, characterized in that, The second unit is also used for: Multimodal feature extraction is performed on the physiological parameter information. Temporal morphological features are extracted using a deep convolutional network, and frequency domain features are extracted using a fast Fourier transform and a spectrum encoder. Nonlinear dynamic features are extracted by calculating sample entropy and Lyapunov exponent. The deep convolutional network uses batch normalization and ReLU activation function, and the spectrum encoder uses layer normalization. A bidirectional attention-enhanced long short-term memory network is used to model the temporal dependency of multimodal features. The long short-term memory network includes a bidirectional long short-term memory layer for forward and backward information fusion, a multi-head self-attention layer for establishing temporal correlation weights, and a residual connection layer for preserving the original features. A multi-scale trend prediction framework is constructed based on the results of temporal dependency modeling, including a short-term predictor based on gated recurrent unit network for five-minute scale prediction, a medium-term predictor based on Transformer encoder for thirty-minute scale prediction, and a long-term predictor based on temporal convolutional network for four-hour scale prediction. The multi-scale trend prediction results are input into the cardiovascular system scoring unit and the respiratory system scoring unit based on the gated attention mechanism to score physiological indicators. The cardiovascular system scoring unit calculates the score based on heart rate variability, blood pressure fluctuation characteristics and body surface electrical signal morphology characteristics. The respiratory system scoring unit calculates the score based on respiratory rate changes, blood oxygen saturation trends and respiratory rhythm characteristics. Trend analysis was performed on the physiological index scores. The trend analysis used an ensemble learning strategy to fuse the prediction results of the autoregressive moving average model and the Prophet model, and the confidence interval of the prediction results was calculated using the Monte Carlo sampling method. A physiological function status report is generated based on the confidence intervals of the physiological indicator scores, trend analysis, and prediction results. The physiological function status report includes cardiovascular system stability analysis results, respiratory system dynamic characteristics, autonomic nervous system regulation capacity indicators, organ function status quantification results, and physiological rhythm homeostasis evaluation.

3. The system according to claim 1, characterized in that, The second unit is also used for: The indicators in the physiological function status report are standardized and dimensionality-reduced preprocessing is performed, and the drug administration environment state vector is formed by combining historical drug administration records and physiological indicator prediction trends. Collect physiological response data to update the drug administration environment state vector, calculate the actual comprehensive reward function value and store it in the experience replay pool, sample data from the experience replay pool to update the strategy network, and when a change in physiological state is detected to exceed a preset change threshold, trigger the drug administration plan adjustment and re-execute the plan safety assurance mechanism to generate the final personalized drug administration plan.

4. The system according to claim 3, characterized in that, By combining the drug action mechanism knowledge base and the drug administration environment state vector, two dual-depth Q-networks with identical structures and independent parameters are constructed. One network is used to select the dosing regimen, and the other network is used to evaluate the Q value of the selected regimen. A comprehensive reward function is designed, which includes physiological indicator improvement reward, drug interaction penalty and dosing safety constraint. The weight coefficients of the comprehensive reward function are determined by grid search optimization. The steps for generating an initial dosing regimen based on a trained deep Q-network include: Two identical but parameter-independent dual-depth Q-networks are constructed. Each dual-depth Q-network includes an input layer, three fully connected layers, a batch normalization layer, and a random deactivation layer. The input layer corresponds to the state vector dimension, and the output layer corresponds to the action space dimension. The first deep Q-network is responsible for action selection, generating a set of candidate actions with the highest Q-values ​​and calculating the Q-value confidence interval for each candidate action. The final action is selected based on an upper confidence bound strategy. The second deep Q-network is responsible for independently evaluating the selected action and calculating the Q-value. The evaluation results of the two networks are used to calculate the Q-value difference. When the Q-value difference exceeds a preset threshold, an uncertainty warning is triggered. The final action evaluation value is obtained by dynamically weighting the Q-values ​​of the two networks. The dynamic weighting coefficients are determined based on the historical prediction accuracy of each network. A comprehensive reward function is constructed, comprising rewards for improving physiological indicators, penalties for adverse reactions when multiple drugs are used simultaneously, dosing safety constraints to ensure that the dosage is within a safe range, rewards for temporal rationality to ensure the rationality of the time interval between adjacent dosings, and rewards for smoothness of dosing dosage. The weight coefficients of the comprehensive reward function are determined using a Bayesian optimization method. The Bayesian optimization method includes establishing a Gaussian process regression model to predict the objective function value of the weight coefficient combination, selecting the next set of weight coefficient combinations to be evaluated based on the maximization expectation improvement criterion, and iteratively optimizing until convergence. The objective function comprehensively considers the accuracy of the dosing regimen, the incidence of adverse reactions, and the stability index of the regimen. The dual-depth Q-network is trained based on a priority experience replay mechanism, and supervised pre-training is performed using expert knowledge during the training process. An initial dosing regimen is generated based on the trained dual-depth Q-network.

5. The system according to claim 1, characterized in that, The third unit is also used for: Energy is harvested by a flexible photovoltaic film on the surface of an intelligent infusion pump in the field. The energy harvested by the flexible photovoltaic film is then input into a microcapsule phase change energy storage material with octadecyl n-alkane as the core for storage. The microcapsule phase change energy storage material has a wall material with controllable cross-linking degree, which is used for phase change energy storage and provides a constant temperature environment for drug preservation. Based on the real-time photovoltaic power of the flexible photovoltaic film, the energy storage power of the microcapsule phase change energy storage material, the real-time temperature of the drug, and the target temperature, a multi-objective optimization function is constructed. The multi-objective optimization function includes an energy balance objective function, a temperature control objective function, and a power consumption allocation objective function. An improved NSGA-III algorithm with an adaptive crossover operator is used to solve the multi-objective optimization function. The crossover probability of the adaptive crossover operator is related to the fitness value of the current solution, and an adaptive update mechanism for reference points and a local search strategy are introduced. Based on the solution results of the multi-objective optimization function, the power consumption of the injection pump system is dynamically allocated in a hierarchical manner according to the priority order of the core functional layer, the auxiliary functional layer and the emergency backup layer. The core functional layer includes a micro peristaltic pump mechanism, a multi-sensor array and a control processor, and the auxiliary functional layer includes a temperature control module and a communication module. Based on weather conditions, the system predicts future light intensity and calculates available power. A fuzzy PID controller is constructed by combining light intensity change rate, battery state of charge, and drug temperature deviation. This controller outputs power adjustment values ​​for each functional module, which includes a micro-peristaltic pump mechanism, a multi-sensor array, a control processor, a temperature control module, and a communication module. A weighted average of the current target power and historical power is used to obtain a transitional power value. Power is allocated to each functional module based on this transitional power value. The system monitors the power utilization rate and actual power value of each functional module in real time. Power reallocation is triggered when the difference between the maximum and minimum power utilization rate exceeds a preset adjustment threshold, or when the deviation between the actual power value and the target power value exceeds a preset range.

6. The system according to claim 5, characterized in that, The improved NSGA-III algorithm with an adaptive crossover operator is used to solve the multi-objective optimization function. The crossover probability of the adaptive crossover operator is related to the fitness value of the current solution. The steps of introducing an adaptive update mechanism for reference points and a local search strategy include: Improvements to the NSGA-III algorithm include the introduction of an adaptive crossover operator, an adaptive update mechanism for reference points, and a local search strategy; A calculation formula for an adaptive crossover operator is constructed, which includes the baseline crossover probability, the normalized fitness value of the current solution, the maximum fitness value of the population, the current iteration number, and the adaptive adjustment coefficient. The crossover probability is dynamically adjusted according to the calculation formula. The crossover probability is inversely proportional to the fitness value of the current solution and inversely proportional to the iteration number. Based on the crossover operation results of the adaptive crossover operator, the reference point is dynamically updated using an exponential moving average method. The position vector of the reference point, the centroid of the non-dominated solution set, and the learning rate are used as parameters to calculate the new reference point position. The distance and average distance between adjacent reference points are calculated, and a distance-based density penalty factor is constructed. When the density penalty factor is greater than the reference point splitting threshold, adjacent reference points are selected in the high-density region to calculate the midpoint position and generate a new reference point. Based on the reference point update results, the initial search radius, the current generation, the maximum generation, and the fitness value are used as parameters to construct a dynamic adjustment formula for the search radius. Based on the gradient information of the current solution, a probability distribution of the search direction is constructed to obtain the local search direction. The population diversity index is calculated based on the local search direction. The population diversity index is compared with the local search trigger threshold and the trigger state of the local search is determined by combining the iteration interval. The local search operation is performed according to the trigger state to obtain the local search result. The local search result is combined with the population high-quality solution retention mechanism to design a population re-initialization strategy. The optimized population is obtained by retaining high-quality solutions, random initialization and local search. Monitor the stability status of the injection pump system, adaptively adjust the optimization period based on the stability status of the injection pump system, and update the optimized population based on the adjusted optimization period.

7. The system according to claim 1, wherein the third unit is further configured to: Real-time data from the pressure sensor, bubble sensor, and temperature sensor in the intelligent injection module are collected. The real-time data is input into a dual-stream attention network for feature extraction. Wavelet transform is used to decompose the pressure data into multiple scales. A target detection network is used to identify and locate the bubble data. Temperature data features are obtained through adaptive piecewise fitting to obtain multimodal sensing features. The multimodal sensing features are input into a cross-modal self-attention network to calculate the correlation weights between features and perform feature fusion. A variational autoencoder is used to learn the sample distribution features, and an anomaly score is calculated by combining the isolated forest algorithm to generate anomaly warning information for the status of the injection pump tubing. Real-time physiological parameters are constructed into a graph structure network, and associated features are extracted through graph convolution operations. A multi-head recursive attention mechanism is introduced to obtain temporal change features. The anomaly warning information is fused with the temporal change features to generate hierarchical safety features. Based on the hierarchical security features, an adaptive fuzzy inference system is constructed. The particle swarm optimization algorithm is used to dynamically optimize the fuzzy rules, and an adversarial training strategy is introduced. The current state is inferred and calculated according to the optimized fuzzy rules to obtain a security risk score, and the security risk score is divided into multiple security risk levels. Based on security risk levels, a hierarchical reinforcement learning framework is used for protection decisions. The higher-level network maps different risk levels to corresponding protection measures, while the lower-level network converts the protection measures into execution instructions. An adaptive PID controller is constructed based on the execution instructions, and a model predictive control algorithm is used to adjust the PID parameters in real time to output protection control instructions of the corresponding level.

Images