Infection risk intelligent monitoring and control system based on machine learning

Abstract

The invention discloses an intelligent infection risk monitoring and control system based on machine learning, and the system comprises a data collection and preprocessing module which is used for collecting multi-source environment data, and generating a multi-dimensional time series data set; the hierarchical Dirichlet working condition identification module is used for establishing a hierarchical Dirichlet working condition model and forming an infection risk working condition identification result; the dynamic control chart construction module is used for constructing a dynamic control chart and forming a dynamic topological structure; the multi-component coupling predictive control module is used for executing predictive control operation and outputting a control execution result; the self-evolution memory potential field control module is used for calculating potential field energy and updating a control target weight; and the optimal control and feedback module is used for generating and executing an optimal control instruction and feeding back an execution result. By fusing hierarchical Dirichlet identification, dynamic prediction control and self-evolution potential field adjustment, real-time monitoring and self-adaptive control of infection risks are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial process control and intelligent monitoring technology, and in particular to an intelligent monitoring and control system for infection risk based on machine learning. Background Technology

[0002] In the fields of public health and industrial environmental safety, the monitoring and control of infection risks has always been a crucial aspect of ensuring environmental safety and human health. Existing infection risk monitoring systems typically rely on fixed sensors and static threshold strategies. By collecting environmental parameters such as air quality, microbial concentration, temperature, and humidity, they determine the infection risk level and trigger ventilation or disinfection measures. While traditional methods offer a degree of real-time capability, they largely depend on manually set rules and lack the ability to adaptively learn and dynamically decide in response to complex environmental changes. They struggle to cope with nonlinear environmental conditions involving multiple coupled factors, and control strategies often lag or over-respond.

[0003] With the development of machine learning and intelligent control technologies, some studies have attempted to combine predictive models with fuzzy control, reinforcement learning, or neural networks for environmental infection risk prediction and intelligent regulation. However, traditional methods generally have two limitations: first, the division of operating conditions relies on a fixed model structure or a preset number of clusters, making it impossible to adaptively adjust the model complexity when environmental characteristics change dynamically; second, control decisions usually adopt single-objective or static weight optimization, failing to achieve dynamic coordination between control target weights, constraint ranges, and environmental states, resulting in difficulty in balancing control stability and energy efficiency in long-term operation.

[0004] Therefore, how to provide an intelligent monitoring and control system for infection risk based on machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent monitoring and control system for infection risk based on machine learning. This invention comprehensively utilizes nonparametric Bayesian hierarchical modeling, dynamic control chart predictive control, and self-evolving memory potential field adjustment technology to construct an adaptive learning intelligent monitoring and control system for infection risk. First, this invention achieves automatic classification of environmental conditions and identification of risk states through hierarchical Dirichlet condition identification. Then, based on dynamic control charts, it executes multi-component coupled predictive control to achieve adaptive adjustment of control objectives and constraint ranges under different conditions. Finally, it uses a self-evolving memory potential field mechanism to dynamically update the control objective weights and constraint parameters, ensuring the system's stability and continuous optimization capability during long-term operation. This invention can achieve real-time monitoring, dynamic response, and self-learning evolution of infection risk in complex multi-source environments, possessing advantages such as high identification accuracy, fast control response, strong robustness, and long-term stable operation.

[0006] An intelligent monitoring and control system for infection risk based on machine learning according to an embodiment of the present invention includes the following modules: The data acquisition and preprocessing module is used to collect multi-source environmental data within the monitoring area, and to preprocess the data to generate a multi-dimensional time-series dataset. The hierarchical Dirichlet working condition identification module is used to establish a hierarchical Dirichlet working condition identification model, generate local working condition distribution and global hyperdistribution, and form infection risk working condition identification results. The dynamic control chart construction module is used to construct dynamic control charts, record risk transfer probabilities and control dependencies, and form a dynamic topology structure to describe the evolution of operating conditions. The multi-component coupled predictive control module is used to perform multi-component coupled predictive control operations, dynamically adjust the constraint set, generate a variable constraint domain, and output the control execution results. The self-evolving memory potential field control module is used to construct the self-evolving memory potential field control module, calculate the potential field energy and generate the state matrix, and dynamically update the control target weights according to the changes in the potential field. The optimal control and feedback module is used to generate optimal control commands and transmit them to the execution unit for real-time adjustment, and to provide feedback on the execution results and environmental monitoring data.

[0007] Optionally, modules can be integrated using the following methods: Collect multi-source environmental data within the monitoring area, preprocess the multi-source environmental data, and generate a multi-dimensional time-series dataset; Based on a multidimensional time series dataset, a hierarchical Dirichlet work condition identification model is established. Non-parametric Bayesian hierarchical modeling is used to generate local work condition distribution and global hyperdistribution. The infection risk work conditions in different time periods and spatial regions are automatically divided and merged. The posterior weights of work condition segments, risk feature parameters and work condition transition relationships are output to form the infection risk work condition identification results. Based on the identification results of infection risk conditions, a dynamic control graph is constructed to determine the infection risk condition nodes and condition transition edges, record the risk transition probability and control dependency relationship between each condition segment, and use the topological structure of the dynamic control graph to describe the evolution state of infection risk conditions. Based on the dynamic control chart, multi-component coupled predictive control operations are performed. The posterior weights and risk characteristic parameters of the operating conditions are used as inputs to determine the control objectives, constraint ranges and control objective weights corresponding to each operating condition. A variable constraint domain is formed by dynamically adjusting the constraint set, and control execution results are generated. A self-evolving memory potential field control module is constructed. The control execution results and risk characteristic parameters are input, a multi-time domain potential field state is established using a time decay function, the potential field energy of each working condition is calculated, and a potential field state matrix is ​​generated. The control target weights are dynamically updated according to the changes in potential field energy. Based on the updated control target weights, the optimal control command is generated and transmitted to the execution unit. The execution results and environmental monitoring data are fed back to the hierarchical Dirichlet condition identification model and the self-evolving memory potential field control module in real time, and the parameters and control target weights are dynamically updated.

[0008] Optionally, the multi-source environmental data includes airborne microbial concentration, particulate matter concentration, temperature and humidity, carbon dioxide concentration, wind speed, population density, and equipment operating status.

[0009] Optionally, the preprocessing of multi-source environmental data includes time synchronization, outlier removal, noise filtering, missing data completion, and normalization.

[0010] Optionally, the formation of infection risk condition identification results includes: The multidimensional time series dataset is organized into windows according to a unified time base and spatial region index. The observation vector composition and feature field set of each time window are determined to form the input sequence. In the hierarchical Dirichlet working condition identification model, nonparametric Bayesian hierarchical modeling is adopted to set the hierarchical relationship between the global hyperdistribution and the local working condition distribution in each region, and to establish a three-module architecture consisting of a prototype dynamic generation module, a spatiotemporal consistency coding module, and a topology-preserving convergence module. In the prototype dynamic generation module, based on the non-parametric prior-driven scalable cluster allocation mechanism, the working condition assignment is completed window by window for the input sequence: When the posterior confidence of the existing prototype for the current window is insufficient and the novelty evaluation reaches the addition threshold, a new working condition prototype is automatically generated. When the separability evaluation of two or more prototypes is lower than the merging threshold, prototype merging is performed, and the output is a set of local working condition prototypes that are adaptively expanded or compressed as the data progresses, along with the corresponding risk characteristic parameters. In the spatiotemporal consistency coding module, domain alignment and dimensional correction are performed on data from different sensor sources and different spatial regions. The temporal stability score and spatial consistency score of each working condition prototype are calculated and used as reliability weights to calibrate the risk characteristic parameters, and the local working condition distribution with reliability weights is output. In the topology-preserving aggregation module, the local working condition distributions of each region are aggregated upwards into a global superdistribution based on nonparametric Bayesian hierarchical modeling. The number and sparsity of working condition segments are adaptively adjusted by the concentration control parameter to maintain the orderliness of risk gradients and topological sparsity constraints between working conditions. The posterior weights of working condition segments and calibrated risk feature parameters are output, and the survival rate and occurrence frequency of working condition segments in each time window and spatial region are recorded. Based on the work condition segment assignment results and survival records of adjacent time windows, the work condition transition relationship is statistically analyzed and normalized. The posterior weight of the work condition segment, the risk feature parameters and the work condition transition relationship are combined to output the infection risk work condition identification result.

[0011] Optionally, the description of the evolution of infection risk conditions using the topology of a dynamic control graph includes: Read the infection risk working condition identification results to form a set of working condition segments, corresponding working condition segment posterior weights, risk characteristic parameters, and working condition transition relationships. Align and index the above information according to a unified time base and spatial region, and use it as input data for dynamic control charts. Based on the input data, nodes and initial connections are established in the dynamic control chart. Each node corresponds one-to-one with a working condition segment, and the connections are determined by a three-channel fusion method. One is the transfer channel weight obtained from the working condition transfer relationship; Second, the neighboring channel weight is based on the similarity of risk characteristic parameters; Third, based on the survivability and reliability channel weights, the three channel weights are boundedly fused through a gating fusion device, and threshold pruning and K-nearest neighbor retention are implemented, while global backoff connections are set. The control dependency of nodes and the importance score of edges are calculated on the dynamic control graph. The control dependency is obtained by weight transfer and normalization in a finite number of rounds on the dynamic control graph. The weighted topology structure with control dependency and edge importance labels is output by combining the posterior weight of the node's working condition segment, the fusion weight of adjacent edges and node metric information. The weighted topology is mapped to control reference values, generating control target weight reference values ​​and constraint mapping reference values ​​for each working condition segment. The control target weight reference values ​​increase monotonically with control dependence, while the constraint mapping reference values ​​are adjusted segmentally according to the importance score of the edge. The dynamic control graph is updated in a timely manner and anomaly is suppressed. A smooth update strategy with a sliding window and an upper limit constraint on the change range are used to update the node attributes and edge weights synchronously with a limited range. Sudden anomalies are truncated and rolled back. Newly emerging working condition segments are cold-started and connected according to a conservative edge connection strategy. The output is a dynamic control graph after time smoothing and anomaly suppression.

[0012] Optionally, the generation of control execution results includes: Read the posterior weights and risk characteristic parameters of the operating condition segment, as well as the reference values ​​of the control target weights, constraint mapping reference values, and weighted topology information, as inputs and operating configurations for predictive control in this cycle; Based on the reference value of the control target weight, the posterior weight of the working condition segment and the availability index corresponding to the equipment operating status, the initial control target weight of each working condition segment is calculated by using an adaptive weight fusion mechanism. The initial control target weight is then normalized and subject to boundary constraints to obtain the control target weight for this period. Based on the constraint mapping reference value, and combined with the edge importance score and risk characteristic parameters of the weighted topology, the constraint range of each working condition segment is incrementally contracted or relaxed, and the minimum safety boundary, slope limit and saturation limit are superimposed to generate a variable constraint domain and mapping parameter set. Control candidates are generated in parallel using a two-domain rolling solution mechanism within the variable constraint domain: Short-period channels generate fast candidate control sequences for safety-related quantities; Long-cycle channels generate economical candidate control sequences for energy consumption and comfort. Based on the control target weights for this period, the candidate control sequences of the two channels are weighted and synthesized to form a joint candidate control sequence; The joint candidate control sequence is subjected to feasibility verification and soft handover processing. When infeasibility or out-of-bounds risk is detected, the constraint mapping parameters and control target weights are adjusted to a limited extent according to the priority-based relaxation strategy. The upper limit of the change range and the dwell time constraint are applied to obtain a smooth handover sequence and output the control execution result at the current time.

[0013] Optionally, the step of calculating the potential energy for each operating condition segment, generating a potential state matrix, and dynamically updating the control target weights based on changes in the potential energy includes: Receive the control execution result, the current control target weight and constraint mapping parameters, set the potential field time window and time decay function, and initialize the self-evolving memory potential field control module composed of multi-time domain potential field construction unit, self-calibration steady-state unit and weight-constraint collaborative update unit; In the multi-time domain potential field construction unit, the risk deviation measure of each working condition is calculated based on the risk characteristic parameters and the current observation data. The risk deviation measure is accumulated and updated in the potential field time window according to the time decay principle to generate the potential field energy of each working condition and aggregate into a potential field state matrix. In the self-calibrating steady-state unit, the survivability and occurrence frequency of the operating condition are used to perform reliability weighted calibration on the potential field state matrix. Combined with the feasibility judgment results of the control execution results and historical out-of-bounds record information, the risk deviation metric is smoothed, limited, and anomaly truncation is performed, and the calibrated and drift-free potential field state matrix is ​​output. In the weight-constraint collaborative update unit, the control target weights and constraint parameters of each working condition are synchronously adjusted according to the calibrated potential field state matrix. When the potential field energy increases, the control target weights are increased and the constraints are tightened. When the energy decreases, the control target weights are decreased and the constraints are relaxed. The updated control target weights and constraint mapping parameters are output.

[0014] Optionally, the step of generating optimal control instructions and transmitting them to the execution unit based on the updated control target weights includes: Read the updated control target weights and constraint mapping parameters, and combine them with the current process state and measurement output to determine the optimal control command at this moment under the constraints. The optimal control command is allocated and sent to the execution unit according to the execution channel. The execution unit includes a fan, a filtration device, a disinfection module and a ventilation device. The slope and amplitude limits are implemented for the optimal control command sending process, and the execution confirmation mark and timestamp are recorded. The system collects execution results and environmental monitoring data, forms a feedback dataset, and sends it back to the hierarchical Dirichlet operating condition identification model and the self-evolving memory potential field control module for dynamic updates of parameters and control target weights. At the same time, it archives the current cycle log as input for the next cycle.

[0015] The beneficial effects of this invention are: This invention introduces a hierarchical Dirichlet working condition identification mechanism to achieve high-precision dynamic modeling and adaptive risk identification of multi-source environmental data. Compared to traditional monitoring methods that rely on fixed thresholds or preset cluster numbers, this invention can autonomously generate, merge, and adjust working condition distributions under varying environmental characteristics, spatial regions, and temporal distributions, accurately identifying the dynamic evolution characteristics of different risk levels. This mechanism enables the system to adaptively identify complex nonlinear environments, significantly improving the sensitivity and real-time performance of infection risk monitoring.

[0016] At the control level, this invention constructs a multi-component coupled predictive control method based on dynamic control charts. By uniformly mapping the posterior weights of operating conditions, risk characteristic parameters, and control dependencies to the dynamic constraint domain, it achieves the co-evolution of control objectives and constraints. The system can dynamically allocate control objective weights and adjust constraint boundaries according to different risk conditions, ensuring both rapid risk response and a balance between energy consumption and stability, overcoming the problems of fixed and delayed response in traditional control strategies. This invention introduces a self-evolving memory potential field control mechanism, combining time decay memory, potential field energy feedback, and parameter self-updating to achieve long-term stable evolution and self-learning optimization of the control system. This mechanism can automatically adjust the control objective weights and constraint parameters based on historical deviations, forming a continuously optimized dynamic equilibrium process, enabling the system to maintain steady-state operation and adaptively optimize over long-term operation. In summary, this invention achieves significant improvements in infection risk identification accuracy, control response speed, and system stability, possessing broad engineering application value and promising prospects for widespread adoption. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the structure of an intelligent monitoring and control system for infection risk based on machine learning proposed in this invention; Figure 2 This is a flowchart illustrating an intelligent monitoring and control method for infection risk based on machine learning proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 A machine learning-based intelligent monitoring and control system for infection risk includes the following modules: The data acquisition and preprocessing module is used to collect multi-source environmental data within the monitoring area, and to preprocess the data to generate a multi-dimensional time-series dataset. The hierarchical Dirichlet working condition identification module is used to establish a hierarchical Dirichlet working condition identification model, generate local working condition distribution and global hyperdistribution, and form infection risk working condition identification results. The dynamic control chart construction module is used to construct dynamic control charts, record risk transfer probabilities and control dependencies, and form a dynamic topology structure to describe the evolution of operating conditions. The multi-component coupled predictive control module is used to perform multi-component coupled predictive control operations, dynamically adjust the constraint set, generate a variable constraint domain, and output the control execution results. The self-evolving memory potential field control module is used to construct the self-evolving memory potential field control module, calculate the potential field energy and generate the state matrix, and dynamically update the control target weights according to the changes in the potential field. The optimal control and feedback module is used to generate optimal control commands and transmit them to the execution unit for real-time adjustment, and to provide feedback on the execution results and environmental monitoring data.

[0020] refer to Figure 2 A machine learning-based intelligent monitoring and control method for infection risk includes: Collect multi-source environmental data within the monitoring area, preprocess the multi-source environmental data, and generate a multi-dimensional time-series dataset; Based on a multidimensional time series dataset, a hierarchical Dirichlet work condition identification model is established. Non-parametric Bayesian hierarchical modeling is used to generate local work condition distribution and global hyperdistribution. The infection risk work conditions in different time periods and spatial regions are automatically divided and merged. The posterior weights of work condition segments, risk feature parameters and work condition transition relationships are output to form the infection risk work condition identification results. Based on the identification results of infection risk conditions, a dynamic control graph is constructed to determine the infection risk condition nodes and condition transition edges, record the risk transition probability and control dependency relationship between each condition segment, and use the topological structure of the dynamic control graph to describe the evolution state of infection risk conditions. Based on the dynamic control chart, multi-component coupled predictive control operations are performed. The posterior weights and risk characteristic parameters of the operating conditions are used as inputs to determine the control objectives, constraint ranges and control objective weights corresponding to each operating condition. A variable constraint domain is formed by dynamically adjusting the constraint set, and control execution results are generated. A self-evolving memory potential field control module is constructed. The control execution results and risk characteristic parameters are input, a multi-time domain potential field state is established using a time decay function, the potential field energy of each working condition is calculated, and a potential field state matrix is ​​generated. The control target weights are dynamically updated according to the changes in potential field energy. Based on the updated control target weights, the optimal control command is generated and transmitted to the execution unit. The execution results and environmental monitoring data are fed back to the hierarchical Dirichlet condition identification model and the self-evolving memory potential field control module in real time, and the parameters and control target weights are dynamically updated.

[0021] In this embodiment, the multi-source environmental data includes airborne microbial concentration, particulate matter concentration, temperature and humidity, carbon dioxide concentration, wind speed, population density, and equipment operating status.

[0022] In this embodiment, the preprocessing of multi-source environmental data includes time synchronization, outlier removal, noise filtering, missing data completion, and normalization.

[0023] In this embodiment, the process of forming the infection risk assessment result includes: The multidimensional time series dataset is organized into windows according to a unified time base and spatial region index. The observation vector composition and feature field set of each time window are determined to form the input sequence. In the hierarchical Dirichlet working condition identification model, nonparametric Bayesian hierarchical modeling is adopted to set the hierarchical relationship between the global hyperdistribution and the local working condition distribution in each region, and to establish a three-module architecture consisting of a prototype dynamic generation module, a spatiotemporal consistency coding module, and a topology-preserving convergence module. In the prototype dynamic generation module, based on the non-parametric prior-driven scalable cluster allocation mechanism, the working condition assignment is completed window by window for the input sequence: When the posterior confidence of the existing prototype for the current window is insufficient and the novelty evaluation reaches the addition threshold, a new working condition prototype is automatically generated. When the separability evaluation of two or more prototypes is lower than the merging threshold, prototype merging is performed, and the output is a set of local working condition prototypes that are adaptively expanded or compressed as the data progresses, along with the corresponding risk characteristic parameters. In the spatiotemporal consistency coding module, domain alignment and dimensional correction are performed on data from different sensor sources and different spatial regions. The temporal stability score and spatial consistency score of each operating condition prototype are calculated and used as reliability weights to calibrate risk characteristic parameters. The resulting local operating condition distribution with reliability weights is output. Specifically, the calculation of the temporal stability score and spatial consistency score for each operating condition prototype involves: The sensor feature vectors of the same prototype under different time windows are statistically analyzed in segments, and the mean and variance are extracted respectively. The statistical results of multiple time periods are compared, the change range of feature distribution between time periods is measured, a stability level standard is constructed, and the scoring level is corresponding to the change range to quantify the stability in the time dimension. The feature representations collected by the same prototype under multiple sensor deployment areas are standardized, and the cosine similarity is used to calculate the similarity of feature vectors between regions. The corresponding scoring level is assigned based on the consistency of spatial distribution. In the topology-preserving aggregation module, nonparametric Bayesian hierarchical modeling is used to aggregate the local load condition distributions of each region into a global superdistribution. The number and sparsity of load condition segments are adaptively adjusted through concentration control parameters to maintain the orderliness of risk gradients and topological sparsity constraints among load conditions. The posterior weights of load condition segments and calibrated risk feature parameters are output. The survival rate and frequency of occurrence of load condition segments in each time window and spatial region are recorded. Specifically, the nonparametric Bayesian hierarchical modeling that aggregates the local load condition distributions of each region into a global superdistribution involves: The labeled or inferred local operating condition distributions in each monitoring area are used as observation variables. Feature statistics, including mean vector, covariance matrix, and category frequency distribution, are extracted for each local distribution to represent the operating condition behavior characteristics of the monitoring area. Based on the extracted regional feature statistics, a nonparametric Bayesian hierarchical structure model is constructed using a Dirichlet process hybrid model. The first layer defines the generation mechanism of the local distribution of each region, and the second layer introduces shared hyperparameters to describe the potential common patterns in the global hierarchical structure. The correlation and convergence patterns between regions are estimated through sampling. After completing the hierarchical modeling, Gibbs sampling was used to iteratively optimize the parameters of the Dirichlet process mixture model to generate a global hyperdistribution that conforms to the statistical characteristics of the observed data. Based on the work condition segment assignment results and survival records of adjacent time windows, the work condition transition relationship is statistically analyzed and normalized. The posterior weight of the work condition segment, the risk feature parameters and the work condition transition relationship are combined to output the infection risk work condition identification result.

[0024] In this embodiment, the description of the evolution of infection risk conditions using the topology of a dynamic control graph includes: Read the infection risk working condition identification results to form a set of working condition segments, corresponding working condition segment posterior weights, risk characteristic parameters, and working condition transition relationships. Align and index the above information according to a unified time base and spatial region, and use it as input data for dynamic control charts. Based on the input data, nodes and initial connections are established in the dynamic control chart. Each node corresponds one-to-one with a working condition segment, and the connections are determined by a three-channel fusion method. One is the transfer channel weight obtained from the working condition transfer relationship; Second, the neighboring channel weight is based on the similarity of risk characteristic parameters; Third, based on the survivability and reliability channel weights, the three channel weights are boundedly fused through a gating fusion device, and threshold pruning and K-nearest neighbor retention are implemented, while global backoff connections are set. The control dependency of nodes and the importance score of edges are calculated on the dynamic control graph. The control dependency is obtained by weight transfer and normalization in a finite number of rounds on the dynamic control graph. The weighted topology structure with control dependency and edge importance labels is output by combining the posterior weight of the node's working condition segment, the fusion weight of adjacent edges and node metric information. The weighted topology is mapped to control reference values, generating control target weight reference values ​​and constraint mapping reference values ​​for each working condition segment. The control target weight reference values ​​increase monotonically with control dependence, while the constraint mapping reference values ​​are adjusted segmentally according to the importance score of the edge. The dynamic control graph is updated in a timely manner and anomaly is suppressed. A smooth update strategy with a sliding window and an upper limit constraint on the change range are used to update the node attributes and edge weights synchronously with a limited range. Sudden anomalies are truncated and rolled back. Newly emerging working condition segments are cold-started and connected according to a conservative edge connection strategy. The output is a dynamic control graph after time smoothing and anomaly suppression.

[0025] In this embodiment, generating the control execution result includes: Read the posterior weights and risk characteristic parameters of the operating condition segment, as well as the reference values ​​of the control target weights, constraint mapping reference values, and weighted topology information, as inputs and operating configurations for predictive control in this cycle; Based on the reference values ​​of control target weights, the posterior weights of operating conditions, and the availability index corresponding to the equipment operating status, an adaptive weight fusion mechanism is used to comprehensively calculate the initial control target weights for each operating condition. The initial control target weights are then normalized and subject to boundary constraints to obtain the control target weights for the current cycle. Specifically, the adaptive weight fusion mechanism is used to comprehensively calculate the initial control target weights for each operating condition. Based on the availability index corresponding to the operating status of each device, a mapping function is constructed to quantify the availability level into the initial priority value of the control target, which reflects the importance ranking of different control targets under the current operating conditions and forms the first weight subset; Using the output posterior weights of the operating conditions, the sensitivity of each control target in the current identification results is evaluated. A dynamic importance correction factor is constructed based on the rate of change and historical frequency to generate a second weight subset. The information entropy of the first and second weight subsets is calculated through an adaptive entropy weight allocation mechanism. Based on the weight dispersion and information gain level of each subset, the optimal fusion ratio coefficient is determined, and the initial control target weights are weighted, superimposed, and normalized. Based on the constraint mapping reference value, and combined with the edge importance score and risk characteristic parameters of the weighted topology, the constraint range of each working condition segment is incrementally contracted or relaxed, and the minimum safety boundary, slope limit and saturation limit are superimposed to generate a variable constraint domain and mapping parameter set. Control candidates are generated in parallel using a two-domain rolling solution mechanism within the variable constraint domain: Short-period channels generate fast candidate control sequences for safety-related quantities; Long-cycle channels generate economical candidate control sequences for energy consumption and comfort. Based on the control target weights for this period, the candidate control sequences of the two channels are weighted and synthesized to form a joint candidate control sequence; The joint candidate control sequence is subjected to feasibility verification and soft handover processing. When infeasibility or out-of-bounds risk is detected, the constraint mapping parameters and control target weights are adjusted to a limited extent according to the priority-based relaxation strategy. The upper limit of the change range and the dwell time constraint are applied to obtain a smooth handover sequence and output the control execution result at the current time.

[0026] In this embodiment, the calculation of potential field energy for each operating condition segment, the generation of a potential field state matrix, and the dynamic updating of the control target weights based on changes in potential field energy include: Receive the control execution result, the current control target weight and constraint mapping parameters, set the potential field time window and time decay function, and initialize the self-evolving memory potential field control module composed of multi-time domain potential field construction unit, self-calibration steady-state unit and weight-constraint collaborative update unit; In the multi-temporal potential field construction unit, the risk deviation metric for each operating condition is calculated based on the risk characteristic parameters and the current observation data. The risk deviation metric is accumulated and updated within the potential field time window according to the time decay principle, generating the potential field energy for each operating condition and converging it into a potential field state matrix. Specifically, the calculation of the risk deviation metric for each operating condition based on the risk characteristic parameters and the current observation data involves: The risk characteristic parameters corresponding to each working condition segment and the observation data collected in real time in the current period are extracted respectively. The residual value corresponding to each dimension is calculated by vector difference method to form the initial vector of risk deviation. A sensitivity scoring matrix is ​​constructed based on the historical fluctuation range of risk characteristics in each dimension. This matrix is ​​used as a weighting coefficient to weight the initial residual vector, resulting in a risk deviation weight vector that reflects the differences in operating condition stability. The weighted and corrected risk deviation weight vector is merged, and a weighted average is used to extract representative indicators to form the risk deviation metric for the working condition section. In the self-calibrating steady-state unit, the survivability and occurrence frequency of the operating condition are used to perform reliability weighted calibration on the potential field state matrix. Combined with the feasibility judgment results of the control execution results and historical out-of-bounds record information, the risk deviation metric is smoothed, limited, and anomaly truncation is performed, and the calibrated and drift-free potential field state matrix is ​​output. In the weight-constraint collaborative update unit, the control target weights and constraint parameters of each working condition are synchronously adjusted according to the calibrated potential field state matrix. When the potential field energy increases, the control target weights are increased and the constraints are tightened. When the energy decreases, the control target weights are decreased and the constraints are relaxed. The updated control target weights and constraint mapping parameters are output.

[0027] In this embodiment, the step of generating optimal control instructions based on the updated control target weights and transmitting them to the execution unit includes: Read the updated control target weights and constraint mapping parameters, and combine them with the current process state and measurement output to determine the optimal control command at this moment under the constraints. The optimal control command is allocated and sent to the execution unit according to the execution channel. The execution unit includes a fan, a filtration device, a disinfection module and a ventilation device. The slope and amplitude limits are implemented for the optimal control command sending process, and the execution confirmation mark and timestamp are recorded. The system collects execution results and environmental monitoring data, forms a feedback dataset, and sends it back to the hierarchical Dirichlet operating condition identification model and the self-evolving memory potential field control module for dynamic updates of parameters and control target weights. At the same time, it archives the current cycle log as input for the next cycle.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to the intensive care unit of a top-tier general hospital. This area, approximately 480 square meters, is equipped with an air purification system, intelligent ventilation devices, and an ultraviolet disinfection module. The environment is extremely sensitive to changes in air cleanliness, temperature, humidity, and personnel density. Previously, this area used a traditional threshold-based risk warning system. During periods of high population density, high equipment load, or nighttime temperature and humidity fluctuations, the system was prone to problems such as delayed infection risk assessment, delayed ventilation response, and mismatched disinfection cycles. This resulted in a short-term increase in airborne microbial concentration of approximately 25%–30%, posing a potential risk of infection transmission.

[0029] In the system deployment of this invention, the multi-source environmental data acquisition and preprocessing module first deploys 15 airborne microbial detection nodes, 12 temperature and humidity sensors, 8 CO2 concentration probes, and 6 infrared personnel statistics modules to collect environmental and personnel activity data in real time, with a sampling frequency of 60 times per minute. The system performs anomaly filtering, noise smoothing, and normalization processing on the collected data to generate a multi-dimensional time-series dataset containing airborne microbial concentration, temperature and humidity, crowd density, CO2 level, and filtration equipment status.

[0030] Subsequently, the hierarchical Dirichlet work condition identification module performed nonparametric Bayesian hierarchical modeling on the dataset. During a seven-day experiment, it automatically identified 12 typical infection-risk work conditions, including "high humidity and high pedestrian traffic," "low ventilation load," and "high CO2 accumulation," and output the corresponding posterior weights and risk feature parameters for each condition. The model can automatically expand the work condition distribution when new conditions appear and automatically merge work condition types when similar environments recur, ensuring the identification results have dynamic adaptive capabilities.

[0031] The system further utilizes a dynamic control graph construction module to establish operating condition nodes and risk transfer topologies, mapping the risk transfer relationships of each operating condition segment into a control dependency structure. For example, during peak shift handover periods, the system identifies the "high humidity + high pedestrian flow" operating condition node and the "high CO2 accumulation" operating condition node as having a transfer probability of 0.62, and marks them as high-priority edges in the control topology so that the control algorithm can respond preferentially.

[0032] At the control level, the multi-component coupled predictive control module executes predictive control based on the identified risk characteristics and posterior weights, dynamically adjusting the constraint domain to optimize airflow, filtration velocity, and UV disinfection time in real time. When the system detects that the CO2 concentration increases from 420ppm to 580ppm and the microbial concentration increases by 18% within 10 minutes, the control module automatically increases the fan speed by 15% and extends the filtration cycle time by 1.8 minutes, achieving rapid risk suppression.

[0033] The self-evolving memory potential field control module establishes a time-decaying potential field based on the control execution results and risk deviations, dynamically accumulating historical risk deviations. During 48 hours of continuous monitoring, the system automatically reduced the fan power weight by approximately 12% during low-traffic periods at night, effectively reducing energy consumption fluctuations, and improved control sensitivity during peak periods, resulting in a 22.7% decrease in peak microbial concentration.

[0034] Ultimately, the optimal control and feedback module applies the updated control target weights to real-time execution, linking the fan, filtration equipment, UV disinfection, and ventilation devices for coordinated control. Feedback data shows that the system's average response time after detecting sudden changes in air quality has been reduced from 28 seconds in the traditional system to 7 seconds, and the time for risk recovery to the safe threshold has been reduced from 12 minutes to 3.5 minutes.

[0035] Table 1. Comparison of Infection Risk Monitoring and Control Effects in Hospital ICUs Based on the Method of the Invention As can be seen from the data in Table 1, the intelligent monitoring and control method for infection risk based on machine learning proposed in this invention has achieved significant performance improvement in the hospital intensive care unit (ICU) environment. In terms of response speed, the environmental response time of the system has been shortened from 28.4 seconds of the traditional system to 7.1 seconds, and the response efficiency has been improved by about 75%. This shows that this invention has achieved a rapid linkage response of multiple execution units (such as fans, filtration equipment) through the dynamic control diagram and predictive control mechanism, effectively avoiding the execution lag after the risk signal is recognized. The infection risk recovery time has been shortened from 12 minutes to 3.5 minutes, indicating that after the system detects an increase in risk, it can complete the combined adjustment of wind speed, ventilation volume, and disinfection strategy faster, and quickly return the infection risk to the safety threshold.

[0036] In terms of environmental control effect, the average concentration of airborne microorganisms has been reduced from 245 CFU / m³ to 183 CFU / m³, a decrease of 25.3%, and the peak value of air CO2 concentration has dropped from 620 ppm to 495 ppm, a reduction of about 20%. This shows that the hierarchical Dirichlet working condition recognition model of this invention can accurately identify high-risk working conditions and actively optimize control strategies, making the operation of ventilation and purification equipment more targeted. The fluctuation range of temperature and humidity has been reduced from ±7.6%RH to ±3.2%RH, a decrease of nearly 58%, reflecting that the system has stronger steady-state maintenance ability in a complex environment and can effectively avoid the re-increase of risks caused by environmental imbalance.

[0037] In terms of energy efficiency and operation stability, the daily average energy consumption of the system has dropped from 52.3 kWh to 42.8 kWh, and the energy saving rate has reached 18.2%. At the same time, the average load rate of the fan has been reduced from 85% to 73%, and the operation is more stable. The comprehensive infection risk index has dropped from 0.71 to 0.49, and the overall risk level has been reduced by 31.4%. The stability coefficient of the control system has been increased from 0.82 to 1.00, an increase of about 22%. This invention not only optimizes the sensitivity and accuracy of risk control, but also improves the long-term operation stability and energy utilization efficiency of the system. In summary, this invention is significantly superior to the traditional method in terms of infection risk response speed, environmental control accuracy, and system stability, fully verifying its technical advantages of achieving intelligence, low energy consumption, and high reliability in a complex medical environment. The above is only the preferred specific implementation manner of this invention, but the protection scope of this invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by this invention, according to the technical solution and inventive concept of this invention, making equivalent substitutions or changes, should be covered by the protection scope of this invention.

Claims

1. A machine learning-based intelligent monitoring and control system for infection risk, characterized in that, include: The data acquisition and preprocessing module is used to collect multi-source environmental data within the monitoring area, perform preprocessing, and generate a multi-dimensional time-series dataset. The hierarchical Dirichlet working condition identification module is used to establish a hierarchical Dirichlet working condition identification model, generate local working condition distribution and global hyperdistribution, and form infection risk working condition identification results. The dynamic control chart construction module is used to construct dynamic control charts, record risk transfer probabilities and control dependencies, and form a dynamic topology structure to describe the evolution of operating conditions. The multi-component coupled predictive control module is used to perform multi-component coupled predictive control operations, dynamically adjust the constraint set, generate a variable constraint domain, and output the control execution results. The self-evolving memory potential field control module is used to construct the self-evolving memory potential field control module, calculate the potential field energy and generate the state matrix, and dynamically update the control target weights according to the changes in the potential field. The optimal control and feedback module is used to generate optimal control commands and transmit them to the execution unit for real-time adjustment, and to provide feedback on the execution results and environmental monitoring data.

2. A machine learning-based intelligent monitoring and control method for infection risk, applied to the machine learning-based intelligent monitoring and control system for infection risk as described in claim 1, characterized in that, include: Collect multi-source environmental data within the monitoring area, preprocess the multi-source environmental data, and generate a multi-dimensional time-series dataset; Based on a multidimensional time series dataset, a hierarchical Dirichlet work condition identification model is established. Non-parametric Bayesian hierarchical modeling is used to generate local work condition distribution and global hyperdistribution. The infection risk work conditions in different time periods and spatial regions are automatically divided and merged. The posterior weights of work condition segments, risk feature parameters and work condition transition relationships are output to form the infection risk work condition identification results. Based on the identification results of infection risk conditions, a dynamic control graph is constructed to determine the infection risk condition nodes and condition transition edges, record the risk transition probability and control dependency relationship between each condition segment, and use the topological structure of the dynamic control graph to describe the evolution state of infection risk conditions. Based on the dynamic control chart, multi-component coupled predictive control operations are performed. The posterior weights and risk characteristic parameters of the operating conditions are used as inputs to determine the control objectives, constraint ranges and control objective weights corresponding to each operating condition. A variable constraint domain is formed by dynamically adjusting the constraint set, and control execution results are generated. A self-evolving memory potential field control module is constructed. The control execution results and risk characteristic parameters are input, a multi-time domain potential field state is established using a time decay function, the potential field energy of each working condition is calculated, and a potential field state matrix is ​​generated. The control target weights are dynamically updated according to the changes in potential field energy. Based on the updated control target weights, the optimal control command is generated and transmitted to the execution unit. The execution results and environmental monitoring data are fed back to the hierarchical Dirichlet condition identification model and the self-evolving memory potential field control module in real time, and the parameters and control target weights are dynamically updated.

3. The intelligent monitoring and control method for infection risk based on machine learning according to claim 2, characterized in that, The multi-source environmental data includes airborne microbial concentration, particulate matter concentration, temperature and humidity, carbon dioxide concentration, wind speed, population density, and equipment operating status.

4. The intelligent monitoring and control method for infection risk based on machine learning according to claim 2, characterized in that, The preprocessing of multi-source environmental data includes time synchronization, outlier removal, noise filtering, missing data completion, and normalization.

5. The intelligent monitoring and control method for infection risk based on machine learning according to claim 2, characterized in that, The results of the infection risk identification include: The multidimensional time series dataset is organized into windows according to a unified time base and spatial region index. The observation vector composition and feature field set of each time window are determined to form the input sequence. In the hierarchical Dirichlet working condition identification model, nonparametric Bayesian hierarchical modeling is adopted to set the hierarchical relationship between the global hyperdistribution and the local working condition distribution in each region, and to establish a three-module architecture consisting of a prototype dynamic generation module, a spatiotemporal consistency coding module, and a topology-preserving convergence module. In the prototype dynamic generation module, based on the non-parametric prior-driven scalable cluster allocation mechanism, the working condition assignment is completed window by window for the input sequence: When the posterior confidence of the existing prototype for the current window is insufficient and the novelty evaluation reaches the addition threshold, a new working condition prototype is automatically generated. When the separability evaluation of two or more prototypes is lower than the merging threshold, prototype merging is performed, and the output is a set of local working condition prototypes that are adaptively expanded or compressed as the data progresses, along with the corresponding risk characteristic parameters. In the spatiotemporal consistency coding module, domain alignment and dimensional correction are performed on data from different sensor sources and different spatial regions. The temporal stability score and spatial consistency score of each working condition prototype are calculated and used as reliability weights to calibrate the risk characteristic parameters, and the local working condition distribution with reliability weights is output. In the topology-preserving aggregation module, the local working condition distributions of each region are aggregated upwards into a global superdistribution based on nonparametric Bayesian hierarchical modeling. The number and sparsity of working condition segments are adaptively adjusted by the concentration control parameter to maintain the orderliness of risk gradients and topological sparsity constraints between working conditions. The posterior weights of working condition segments and calibrated risk feature parameters are output, and the survival rate and occurrence frequency of working condition segments in each time window and spatial region are recorded. Based on the work condition segment assignment results and survival records of adjacent time windows, the work condition transition relationship is statistically analyzed and normalized. The posterior weight of the work condition segment, the risk feature parameters and the work condition transition relationship are combined to output the infection risk work condition identification result.

6. The intelligent monitoring and control method for infection risk based on machine learning according to claim 2, characterized in that, The description of the evolution of infection risk conditions using the topology of a dynamic control graph includes: Read the infection risk working condition identification results to form a set of working condition segments, corresponding working condition segment posterior weights, risk characteristic parameters, and working condition transition relationships. Align and index the above information according to a unified time base and spatial region, and use it as input data for dynamic control charts. Based on the input data, nodes and initial connections are established in the dynamic control chart. Each node corresponds one-to-one with a working condition segment, and the connections are determined by a three-channel fusion method. One is the transfer channel weight obtained from the working condition transfer relationship; Second, the neighboring channel weight is based on the similarity of risk characteristic parameters; Third, based on the survivability and reliability channel weights, the three channel weights are boundedly fused through a gating fusion device, and threshold pruning and K-nearest neighbor retention are implemented, while global backoff connections are set. The control dependency of nodes and the importance score of edges are calculated on the dynamic control graph. The control dependency is obtained by weight transfer and normalization in a finite number of rounds on the dynamic control graph. The weighted topology structure with control dependency and edge importance labels is output by combining the posterior weight of the node's working condition segment, the fusion weight of adjacent edges and node metric information. The weighted topology is mapped to control reference values, generating control target weight reference values ​​and constraint mapping reference values ​​for each working condition segment. The control target weight reference values ​​increase monotonically with control dependence, while the constraint mapping reference values ​​are adjusted segmentally according to the importance score of the edge. The dynamic control graph is updated in a timely manner and anomaly is suppressed. A smooth update strategy with a sliding window and an upper limit constraint on the change range are used to update the node attributes and edge weights synchronously with a limited range. Sudden anomalies are truncated and rolled back. Newly emerging working condition segments are cold-started and connected according to a conservative edge connection strategy. The output is a dynamic control graph after time smoothing and anomaly suppression.

7. The intelligent monitoring and control method for infection risk based on machine learning according to claim 2, characterized in that, The generation of control execution results includes: Read the posterior weights and risk characteristic parameters of the operating condition segment, as well as the reference values ​​of the control target weights, constraint mapping reference values, and weighted topology information, as inputs and operating configurations for predictive control in this cycle; Based on the reference value of the control target weight, the posterior weight of the working condition segment and the availability index corresponding to the equipment operating status, the initial control target weight of each working condition segment is calculated by using an adaptive weight fusion mechanism. The initial control target weight is then normalized and subject to boundary constraints to obtain the control target weight for this period. Based on the constraint mapping reference value, and combined with the edge importance score and risk characteristic parameters of the weighted topology, the constraint range of each working condition segment is incrementally contracted or relaxed, and the minimum safety boundary, slope limit and saturation limit are superimposed to generate a variable constraint domain and mapping parameter set. Control candidates are generated in parallel using a two-domain rolling solution mechanism within the variable constraint domain: Short-period channels generate fast candidate control sequences for safety-related quantities; Long-cycle channels generate economical candidate control sequences for energy consumption and comfort. Based on the control target weights for this period, the candidate control sequences of the two channels are weighted and synthesized to form a joint candidate control sequence; The joint candidate control sequence is subjected to feasibility verification and soft handover processing. When infeasibility or out-of-bounds risk is detected, the constraint mapping parameters and control target weights are adjusted to a limited extent according to the priority-based relaxation strategy. The upper limit of the change range and the dwell time constraint are applied to obtain a smooth handover sequence and output the control execution result at the current time.

8. The intelligent monitoring and control method for infection risk based on machine learning according to claim 2, characterized in that, The calculation of potential field energy for each operating condition segment, the generation of a potential field state matrix, and the dynamic updating of control target weights based on changes in potential field energy include: Receive the control execution result, the current control target weight and constraint mapping parameters, set the potential field time window and time decay function, and initialize the self-evolving memory potential field control module composed of multi-time domain potential field construction unit, self-calibration steady-state unit and weight-constraint collaborative update unit; In the multi-time domain potential field construction unit, the risk deviation measure of each working condition is calculated based on the risk characteristic parameters and the current observation data. The risk deviation measure is accumulated and updated in the potential field time window according to the time decay principle to generate the potential field energy of each working condition and aggregate into a potential field state matrix. In the self-calibrating steady-state unit, the survivability and occurrence frequency of the operating condition are used to perform reliability weighted calibration on the potential field state matrix. Combined with the feasibility judgment results of the control execution results and historical out-of-bounds record information, the risk deviation metric is smoothed, limited, and anomaly truncation is performed, and the calibrated and drift-free potential field state matrix is ​​output. In the weight-constraint collaborative update unit, the control target weights and constraint parameters of each working condition are synchronously adjusted according to the calibrated potential field state matrix. When the potential field energy increases, the control target weights are increased and the constraints are tightened. When the energy decreases, the control target weights are decreased and the constraints are relaxed. The updated control target weights and constraint mapping parameters are output.

9. The intelligent monitoring and control method for infection risk based on machine learning according to claim 2, characterized in that, The step of generating optimal control instructions based on the updated control target weights and transmitting them to the execution unit includes: Read the updated control target weights and constraint mapping parameters, and combine them with the current process state and measurement output to determine the optimal control command at this moment under the constraints. The optimal control command is allocated and sent to the execution unit according to the execution channel. The execution unit includes a fan, a filtration device, a disinfection module and a ventilation device. The slope and amplitude limits are implemented for the optimal control command sending process, and the execution confirmation mark and timestamp are recorded. The system collects execution results and environmental monitoring data, forms a feedback dataset, and sends it back to the hierarchical Dirichlet operating condition identification model and the self-evolving memory potential field control module for dynamic updates of parameters and control target weights. At the same time, it archives the current cycle log as input for the next cycle.

Images